How much should I charge for AI automation services in 2026?

In 2026, AI automation services command premium pricing due to strong market demand and measurable business ROI. Basic chatbot services start at $1,500–$5,000 setup and $500–$1,500/month. Full office automation packages for medium businesses can run $10,000–$35,000 setup with $2,500–$7,500/month retainers. The key is always to anchor your price to client ROI — if your automation saves a client $10,000/month in labor, charging $2,000/month is an easy sell.

What is a typical monthly retainer for an AI agency client?

Typical AI agency monthly retainers in 2026 range from $500/month for a basic single-workflow bot (e.g., a website chatbot for a solopreneur) up to $5,000–$8,000/month for enterprise-grade multi-workflow packages with ongoing optimization and reporting. Most successful AI agencies target $1,500–$3,000/month as a sweet spot for small-to-medium business clients — high enough to generate strong recurring revenue, low enough to be a no-brainer relative to the value delivered.

Should I charge an upfront setup fee plus monthly, or just monthly?

The setup fee + monthly retainer model is the industry standard for good reason: the setup fee covers your time to build and configure the automation, while the retainer covers ongoing maintenance, optimization, and support. This model also increases client commitment (they've invested upfront) and reduces churn. Avoid charging only monthly — it undervalues the significant build time and creates financial pressure if a client churns in month 2 after you've done all the heavy lifting.

What profit margins should I expect as an AI automation agency?

AI automation agencies typically operate at 55–80% net profit margins because the primary cost is your time, with minimal overhead. Your main expenses are: (1) software tools/subscriptions ($100–$500/month for platforms like Make, n8n, Zapier, GoHighLevel, OpenAI API — API usage is the fastest-growing line item; open-weight alternatives like Qwen 3.8 Max at $2/$6 per 1M tokens, Kimi K3 at $3/$15 per 1M tokens, and GLM-5.2 with open weights are the counter-lever — agencies that run them keep more of the margin), (2) time for builds and client calls, and (3) any subcontractors or specialized help. Chatbot and social media automation services tend to have the highest margins (70–82%) because they're templated. Full office automation has lower margins (45–65%) due to higher custom build time.

How do I justify my AI agency pricing to skeptical clients?

The best way to justify AI agency pricing is to convert your service value into dollars. Calculate: (1) Hours saved per month × average hourly cost of the work being automated. For example, if your automation saves 40 hours/month and the equivalent labor costs $25/hour, that's $1,000/month in savings — making a $1,500/month retainer look expensive. But if that 40 hours is a $50/hour admin position ($2,000/month), your pricing looks like a bargain. Always quantify the ROI before the sales conversation.

Should I offer a free trial or discovery call for AI services?

A free 30-minute discovery call (not a free trial of the service) is the industry standard and highly recommended. During the call, you diagnose the client's workflow pain points, identify automation opportunities, and present a clear ROI case before quoting. Free service trials, however, are generally discouraged — they create expectations of free work and attract clients who don't value the service. Instead, offer a "pilot project" at a reduced rate (50–75% of normal) with a 30-day guarantee if you want to reduce friction with skeptical prospects.

What tools do AI agencies use, and how do they affect pricing?

The most common AI agency tool stack in 2026 includes: Make.com ($10–$100+/mo), n8n (self-hosted free or $50+/mo cloud), Zapier ($20–$600+/mo), GoHighLevel ($97–$497/mo), OpenAI API ($5–$200+/mo), Voiceflow or Botpress for chatbots ($20–$100+/mo). Open-weight model APIs (Qwen 3.8 Max at $2/$6 per 1M tokens, Kimi K3 at $3/$15 per 1M tokens, GLM-5.2 with open weights) now offer a lower-cost alternative to premium flagship pricing — so you can cap your tool-stack spend even while frontier API prices firm. Your tool costs should be built into your pricing — either passed through to clients as add-ons or bundled into your retainer (preferred for simplicity). A typical tool stack runs $200–$500/month, which should be factored into your margin calculations.

Do open-weight models like Qwen 3.8 Max, Kimi K3, or GLM-5.2 change what I should charge?

Not the price you charge — the margin you keep. Client pricing is value-based: if your automation saves a client $2,000/month in labor, a $1,500 retainer is justified regardless of the model underneath. What open-weight models change is your cost side. Alibaba's Qwen 3.8 Max (2.4T-parameter MoE, 1M context) prices at $2/$6 per 1M tokens with open weights live on Hugging Face since Aug 12, 2026; Moonshot's Kimi K3 is a 2.8T-parameter open-weight model with a 1M-token context priced at $3/$15 per 1M tokens, and Zhipu's GLM-5.2 ships open weights under MIT — all within striking distance of paid frontier models on vendor-run coding benchmarks, though none matches them everywhere. NEW Aug 26, 2026: Z.ai confirmed the anonymous Ox Alpha model was GLM-5.3-Flash — a 320B/18B multimodal open-weight model at $0.15/$0.50 per 1M tokens (50% promo $0.075/$0.25 through Sep 9, 2026), MIT weights live, below DeepSeek V4 Pro off-peak on both axes. NEW Aug 28, 2026: Tencent Hy4 preview — 770B/49B MoE, Apache 2.0, native 1M context — lists at $0.834/$2.501 per 1M tokens on OpenRouter (cached input $0.042): ~40% below GLM-5.3 on input, ~43% below on output, ~6x below Kimi K3 on output, but still above DeepSeek V4 Pro off-peak ($0.66/$1.98); it is a preview with over-verification tendencies and ~770GB FP8 weights, so treat it as an evaluation option with a fallback, not a single point of failure. Early real-world reports on Qwen 3.8 Max (Aug 6, 2026) are promising on fast, multimodal builds but are self-reported with no independent replication — benchmark on your own workloads before promising anything. Keep workflows model-portable, and treat the cheaper token bill as margin, not as a reason to discount. One caveat — now finalized: the Qwen3.8-Max License requires a separate Qwen license before commercial use only if you run a Model-as-a-Service or AI Work Assistant business AND aggregate revenue exceeds $50M over any consecutive 12 months (no % rate — individually negotiated; internal use exempt). Most agencies never cross that bar, but if you resell hosted Qwen access or ship a coding/office-productivity agent product at scale, budget for that licensing layer.

What is GLM-5.3-Flash (Ox Alpha) and how does its pricing compare to DeepSeek?

GLM-5.3-Flash is Z.ai's newly revealed open-weight model — the anonymous "Ox Alpha" that topped OpenRouter usage from Aug 20, confirmed by Z.ai on Aug 26, 2026. Specs: 320B total / 18B active MoE, natively multimodal (text/image/video), 1M-token context, MIT-licensed weights on Hugging Face, served entirely on Chinese AI chips during the free preview week. Official API pricing is $0.15 per 1M input / $0.50 per 1M output (cached input $0.03), with a 50% launch promo to $0.075/$0.25 through Sep 9, 2026 — roughly one-tenth of the flagship GLM-5.3 rate ($1.40/$4.40) and below DeepSeek V4 Pro off-peak on both axes ($0.66/$1.98). On a 10K-in/2K-out task it is ~4x cheaper than DeepSeek V4 Pro off-peak at list and ~32x cheaper than GPT-5.6 Sol. Caveats: benchmarks (DeepSWE v1.1 63.4, Terminal-Bench 2.1 84.3, AutomationBench 48.8, Z.ai Code Bench 29.0 vs Claude Opus 4.8's 29.5) are vendor-reported with no independent replication yet, and MIT weights do not mean easy self-hosting — 320B/18B needs serious hardware (SGLang/vLLM/TokenSpeed/KTransformers). Select GLM-5.3-Flash in the Model Strategy dropdown to model this stack, and read the full analysis at GLM-5.3-Flash Pricing: Ox Alpha Revealed at $0.15/M.

What is Qwen3.8-Flash and how does its pricing compare to GLM-5.3-Flash and DeepSeek?

Qwen3.8-Flash is Qwen's new open-weight multimodal MoE (released Aug 26, 2026) — 125B total / ~6B active per token, ~95% sparsity, 262,144-token native context expandable to 1M via YaRN, weights on Hugging Face (Qwen/Qwen3.8-Flash-Next) + ModelScope, under the qwen-community-1.0 license (not Apache-2.0); production API went live on QwenCloud the same day. Official API pricing: $0.15 per 1M input / $0.47 per 1M output (cache read $0.016; China ¥0.8/¥2.7/¥0.1), live on OpenRouter at $0.15/$0.47 — below GLM-5.3-Flash at list on output ($0.50) and cache ($0.03), below DeepSeek V4 Pro off-peak on both axes ($0.66/$1.98), and one of the lowest cache-read rates on this page; Artificial Analysis medians are $0.30/$1.20. Caveats: benchmarks are vendor-reported with no independent replication yet, it is a "Next" preview of the Qwen4 architecture (expect rapid iteration), and the qwen-community-1.0 license is a new licensing layer. Select Qwen3.8-Flash in the Model Strategy dropdown to model this stack, and read the full analysis at Qwen3.8-Flash-Next Cost: Open-Weight MoE at $0.15/M.

Is OpenAI batch API still available on OpenRouter?

Not through the :batch model suffixes. On Aug 26, 2026, OpenRouter silently removed all 35 openai/*:batch model slugs (plus z-ai/glm-5.2:batch and moonshotai/kimi-k2.7-code:batch) from its routing table — the slugs still answer metadata requests, so nothing looks broken until a completion returns 404 "No endpoints found for openai/gpt-5:batch." The base models are fine: openai/gpt-5 still routes at $1.25/$10.00 per 1M, and the OpenAI vendor page still lists 96 models. You lost a pricing tier, not a model. The 50% batch discount is still reachable three ways: (1) OpenRouter's beta Batch API (POST /api/beta/batches, plain model slug, 24-hour window, text-only — no image/audio/file parts); (2) the openai/flex service tier via provider: {"only": ["openai/flex"]} — synchronous, exactly half of standard, but GPT-5.6 family only (not gpt-5, gpt-4o, o3 or gpt-4.1); (3) OpenAI's own direct Batch API — still 50% off synchronous, separate rate-limit pool, 50K requests/batch, completes within 24h. Check any slug with curl https://openrouter.ai/api/v1/models/ /endpoints — an empty array means the ID is dead. Full routing guide: OpenRouter OpenAI Batch Endpoints Down: 3 Ways to Keep the 50% Discount.

How much cheaper is DeepSeek V4 Pro on cache reads than Claude Fable 5.1?

For long agent runs that re-read context on every step, the cache-read price is the cost lever that matters — and DeepSeek V4 Pro's is still cheaper than Claude Fable 5.1's, but the gap narrowed sharply on Sept 1, 2026. Claude Fable 5.1 cache reads: $0.25 per 1M tokens (Anthropic official, down 75% from Fable 5's $1.00). DeepSeek V4 Pro cache reads: $0.022 per 1M off-peak (current official pricing) — a ~11x gap. The widely-shared "276x" figure from @JulianGoldieSEO compared Fable 5's $1.00 to DeepSeek's launch cache-read price of $0.003625 — arithmetically correct (275.9x) but stale on both sides after DeepSeek's 2026-08-16 increase and Fable 5.1's cache-read cut (gap now ~11x off-peak / ~6x peak vs Fable 5.1). The 92% cache hit rate Julian cites for agent workloads is an assumption, not a published OpenRouter statistic — plausible because agents re-read the same instructions hundreds of times per task, but benchmark your own workload. Model it here: pick DeepSeek V4 Pro (cache-aware) or Claude Fable 5.1 in the Model Strategy dropdown, or use the cache-aware scenario ↓. Julian's full post →

Is DeepSeek still the cheapest option after its announced API price increase?

The increase landed — and DeepSeek still holds the lowest cache-read rates, now led by DeepSeek V4.1-Flash at $0.003/1M off-peak (deepseek-flash, Sept 10, 2026), with the V4 Pro row second at $0.022/1M. DeepSeek's official peak/off-peak rates took effect 2026-08-16 at 16:00 UTC: off-peak V4 Pro $0.022 cache read / $0.66 cache miss / $1.98 output per 1M tokens (peak: $0.044 / $1.32 / $3.96). The old pre-hike schedule (V4-Flash $0.14/$0.28, V4-Pro $0.435/$0.87) is superseded. Even after the increase, V4 Pro's cache-read price is ~11x below Claude Fable 5.1's $0.25 (Fable 5.1's cache reads dropped 75% from Fable 5's $1.00 on Sept 1, 2026) — still the dominant line item for long agent runs — so a cache-aware DeepSeek stack remains the cheapest option for that workload class, but Anthropic has narrowed the gap. What changed: the previous "276x cheaper" headline (from Julian Goldie's Aug 16 post) compared Fable 5's $1.00 to DeepSeek launch pricing; the current gap vs Fable 5.1 is ~11x off-peak / ~6x peak. Keep workflows portable, verify api-docs.deepseek.com/quick_start/pricing before quoting, and remember the open-weight escape hatch: V4 weights are MIT-licensed, so sustained workloads can be self-hosted (vLLM-class serving).

What is Grok 4.6 and how does its API pricing compare to GPT-5.6 Sol?

Grok 4.6 is SpaceXAI's (formerly xAI) frontier model, released Aug 12, 2026. API pricing is $2.00 per 1M input tokens and $6.00 per 1M output tokens — a fast variant costs 2x, and cache hits run $0.50 per 1M input (-75%). Artificial Analysis reports a 500k-token context window and an Intelligence Index of 61, matching GPT-5.6 Sol (max). Against GPT-5.6 Sol's current $4/$20 per 1M (promo window Aug 21 – Nov 21, 2026; previously $5/$30), Grok 4.6 is 50% cheaper on input tokens and 70% cheaper on output tokens — the dimension that dominates cost in reasoning-heavy agent workloads. It is available today in Cursor, Grok Build, the SpaceXAI API, OpenRouter, Vercel, and Cloudflare, with 2x included usage inside Cursor and Grok Build for the first week. Caveats: the 500k context window is AA-reported (x.ai's announcement does not state it), and x.ai's own chart omits Claude Opus 5 (63) which scores above both Grok 4.6 and GPT-5.6 Sol. Select Grok 4.6 in the Model Strategy dropdown to model this stack. x.ai announcement → · Artificial Analysis →

What is Gemini 3.8 Flash and how does its API pricing compare?

Gemini 3.8 Flash is Google's best reasoning & coding model yet, released Sept 2, 2026 — the third Flash-family release in six weeks (3.6 → 3.7 → 3.8) and Google's recommended model for software engineering, autonomous agents, and complex multi-step reasoning. API intro pricing is $0.75 per 1M input tokens and $3.75 per 1M output tokens (output includes thinking tokens) — the same intro rate as Gemini 3.7 Flash, valid through 2026-12-31, then it reverts to $1.50/$7.50 from Jan 1, 2027. It has a 1M-token context window, 64K max output, and Google says it outperforms 3.7 Flash on DeepSWE v1.1 (long-horizon software engineering) and agentic benchmarks — while noting it can use more tokens at higher effort levels. Rolling out to Google AI Pro/Ultra (Gemini app, AI Mode, Sheets) and developers via Gemini API / AI Studio / Antigravity. Gemini 3.8 Flash Cyber is a gated cybersecurity variant (Fairwind program) — not generally available, so do not price it into client stacks. During the intro window it is the cheapest verified frontier-class coding API on this page — undercutting Qwen 3.8 Max ($2/$6) and Grok 4.6 ($2/$6) on both axes. Select Gemini 3.8 Flash in the Model Strategy dropdown to model this stack. Full pricing analysis: Gemini 3.8 Flash API pricing. Google announcement → · Google blog → · Gemini API pricing → · Model card →

What is Gemini 3.7 Flash (previous version) and how does its API pricing compare?

Gemini 3.7 Flash was Google's workhorse Flash model for coding and agents from Aug 13, 2026 until it was superseded by Gemini 3.8 Flash on Sept 2, 2026 — a refinement of Gemini 3.6 Flash (which launched ~3 weeks earlier). API intro pricing is $0.75 per 1M input tokens and $3.75 per 1M output tokens (output price includes thinking tokens) — valid through 2026-12-31, then it reverts to $1.50/$7.50 from Jan 1, 2027 (its post-intro price equals 3.6 Flash's launch price; the intro rate is exactly half). It has a 1M-token context window, 64K max output, a free tier, 5,000 free search requests/month shared across Gemini 3.x, and a 50% batch discount. vs Gemini 3.6 Flash it is a big step up on coding/agent benchmarks: FrontierCode 1.1 Main 43.6% (vs 34.4%), DeepSWE v1.1 65.3% (vs 49.0%), WebDev Arena Elo 1588 (vs 1538). During the intro window it is the cheapest verified frontier-class API on this page — undercutting Qwen 3.8 Max ($2/$6) and Grok 4.6 ($2/$6) on both axes. Select Gemini 3.7 Flash in the Model Strategy dropdown to model this stack. Google blog → · Gemini API pricing → · Model card →

Is Claude Sonnet 5's $2/$10 pricing permanent or scheduled to increase?

Permanent. Anthropic made Claude Sonnet 5's API pricing permanent on Aug 10, 2026: $2.00 per 1M input tokens and $10.00 per 1M output tokens (cache hit $0.20 per 1M input; cache write $2.50 for 5m / $4 for 1h per 1M). The previously scheduled Sept 1, 2026 increase to $3/$15 was CANCELLED — there is no scheduled price change pending. This is an API-pricing change only; subscription prices are unchanged. Input price sits at open-weight parity (Qwen 3.8 Max $2/$6), while output is well below premium frontier (GPT-5.6 Sol $4/$20 promo). Select Claude Sonnet 5 in the Model Strategy dropdown to model this stack. Anthropic announcement → · Claude pricing docs →

What does the Agent Plugins standard mean for how I price agency work?

Agent Plugins 1.0.0 (published Aug 6, 2026) lets you package Agent Skills and MCP servers into one portable plugin that runs across Codex, ChatGPT, Cursor, GitHub Copilot, Kiro, and VS Code. The old cost model — charging a fixed per-platform adaptation fee and maintaining N client-specific forks — now overstates work for plugin-based deliverables. Apply a portability discount (roughly 60–90% off adaptation cost per additional compatible client), treat plugin packaging & distribution as its own priced deliverable, and keep one maintenance stream instead of N. Because v1.0.0 leaves out permissions, sandboxing, and provenance, budget plugin vetting/signing as a compliance service — it's a sellable differentiator, and it protects you when you buy third-party plugins for client environments. See the Agent Plugins section above ↓

What are Cloudflare Wallets and how do wallet spend caps change my cost forecasts?

Cloudflare Wallets (announced Aug 4, 2026, Agents Week) give AI agents a human-readable wallet handle for paying APIs, MCP tools, content, and AI inference within limits set by the wallet's creator. There are two wallet types: Account Wallets (human/organization-owned; hold stablecoins, delegate spending) and Virtual Wallets (agent-owned, operate via API keys; an agent's maximum spend is capped by the Account Wallet owner). Per Virtual Wallet, the owner can define a spending cap / allowance, an approved merchant allow-list, and a maximum transaction size; agents that hit a limit must request a manual override from a human. Payments run on x402 (HTTP 402 with a price manifest). Fees, stablecoins, and launch timing are still undisclosed — full access arrives "in the coming months." What changes for you: worst-case spend per agent becomes a known number (the cap), so monthly AI spend forecasts get a hard ceiling — that's exactly what the Agent Wallet & Spend Cap Estimator above models, capped vs uncapped. Use it before quoting a retainer: if your agent workloads would exceed the client's wallet cap, either the client's cap needs raising (with a documented manual override) or the retainer needs a variable-compute pass-through.

What is the compute supply scenario on this calculator?

It is a supply-side risk module that compares your baseline per-token cost against a power-constrained outlook for 2026-2027. The driver: Gartner forecasts global data center electricity up 26% in 2026 to 565 TWh, passing 1,200 TWh by 2030 where grid supply may be insufficient — power, not chips, is becoming the binding constraint. The offsets: capacity already locked in years ahead (up to 5GW AWS Trainium under Anthropic's >$100B/10yr commitment, and ~3.5GW Google TPU via Broadcom from 2027). You set a grid/power pass-through percentage (ESTIMATE) and a locked-in capacity relief percentage (ESTIMATE); the scenario shows the monthly and 12-month delta vs your baseline. Both percentages are projections, not published prices, and no capital/capacity/lease terms were disclosed for Theseus Infrastructure. Run the Compute Supply Scenario above ↑

Do data-center bans affect AI capacity and what I pay for tokens?

Yes — and the risk now starts before the API. More than 500 local jurisdictions ban or restrict new data-center construction, up from roughly 300 in late June 2026, and 7 in 10 Americans oppose AI data-center builds in their area. Bans don't move today's API list prices; they constrain the supply of new capacity, raise buildout and energy costs, and stretch lead times — the cost base that sets tomorrow's compute prices. Model the risk with a labeled infrastructure-risk buffer (an illustrative 5-10% add-on on compute line items plus a lead-time buffer in weeks) and re-run the estimate quarterly. Full analysis: Data Center Bans: The New AI Capacity Risk Agencies Face →

How many clients do I need to make $10,000/month as an AI agency?

At an average retainer of $1,500/month, you need 7 clients. At $2,000/month average retainer, just 5 clients. At $3,000/month, only 4 clients. This is why positioning for medium-to-large clients (who can afford $2,000–$5,000/month) dramatically reduces client load while increasing revenue. Niching down into high-ROI industries like real estate, insurance, legal, or medical also lets you charge premium rates because the value of automation in those sectors is especially high.

Is it better to niche down or offer all AI services?

Niche down — at least initially. Specializing in one industry (e.g., "AI automation for real estate agents" or "email automation for e-commerce brands") makes your marketing dramatically more effective, allows you to charge premium rates as a specialist, and lets you build productized service packages that you can deliver faster and at higher margins. Once you have a proven playbook in one niche, expanding to adjacent industries becomes much easier and less risky than starting as a generalist.

When should I raise my rates as an AI agency owner?

You should raise rates when: (1) You're consistently closing 80%+ of prospects — demand exceeds supply of your time. (2) You have 3+ strong case studies showing measurable client ROI. (3) You've been at the same rate for 6+ months. (4) Competitors are charging more for similar work. A simple strategy: raise rates by 20–30% on all new clients, then grandfather existing clients at old rates for 6 months before gradual increases. Many agency owners undercharge for years out of fear — but higher prices often attract better, more committed clients.

What do AI coding agents really cost, beyond the per-seat price?

AI coding agent entry plans run about $10–$20 per seat per month (GitHub Copilot at $10; Cursor, Claude Code, Windsurf, and OpenAI Codex around $20). The real cost is billing model × usage. TrueFoundry's Aug 6, 2026 guide breaks agents into three billing structures — flat per-seat (Cursor Pro, Windsurf Pro, Claude Pro), seat-plus-credits (GitHub Copilot Pro/Pro+/Max), and pay-per-token API (Claude Code API mode, OpenAI Codex) — and SpaceXAI's Grok Bot (Aug 11, 2026) adds a fourth: subscription-bundled agent access with no per-token meter and no standalone price (included from $20/mo with any SuperGrok, Cursor Pro or Cursor Teams plan since Aug 26, 2026, on macOS and iOS). Light autocomplete vs. all-day agent workflows can differ 20–50× in token consumption, so heavy users routinely land in the $100–$400+ range. Full pricing guide →

Are Claude Code weekly usage limits going up or down on September 14, 2026?

Both — it depends which allowance you compare against, and the lower one is the one in force now. Anthropic's temporary +50% weekly Claude Code boost ended September 13, 2026 at 11:59 PM PT , and the permanent level that replaced it on September 14, 2026 is 25% higher than the pre-promotion baseline for Pro, Max, Team, and seat-based Enterprise plans [1][2]. Index the pre-promotion limit at 100: the promotion made it 150, the permanent allowance is 125 — so 125 ÷ 100 = +25% (the raise Anthropic announced) and 125 ÷ 150 = −16.7% (≈ −17%) against the allowance users actually had through Sept 13, which is the cut Anthropic published itself: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." [1][3][4] The change covers weekly limits only : 5-hour session limits, plan pricing, and API billing are unchanged [1][6]. The permanent sentence names Pro, Max, Team, and seat-based Enterprise and does not restate the promotion's exclusions, so no claim is made here about Free or consumption-based Enterprise seats [1]; Anthropic publishes no fixed weekly message or token counts for any plan, so every figure here is an index ratio [1][6][9]. Use the Claude Code Usage Limit Cost Impact Estimator on this page to model where your workload lands on the permanent 125 index. Sources: Anthropic Help Center — Claude Code May–August 2026 weekly limits promotion (updated 2026-09-14) · @ClaudeDevs (Aug 29, 2026) · BleepingComputer · MacObserver · Notebookcheck

How much will the Sept 14 Claude Code weekly limit cut cost my agency?

Use the index, not a guess at absolute numbers: the permanent weekly allowance in force since Sept 14, 2026 is 83.3% of the promotion allowance (125/150), so any workload that used 100% of the promotion allowance now consumes 120% of the permanent 125 index — the same weekly workload needs about 20% more plan headroom or usage credits [1][4]. In dollar terms, holding the promotion-era capacity costs roughly 1.2× the seat line that covered it: a Max 5x seat at $100/mo ≈ $120/mo equivalent on the permanent index (arithmetic illustration, not a published price). Anthropic publishes no fixed weekly message or token counts for any plan, so these are index ratios, not absolute limits [1][6][9]. Run the Claude Code Usage Limit Cost Impact Estimator on this page with your own index basis, usage %, seat price, and headcount. Full analysis: Claude Code Usage Limit Cost Impact .

How much does Grok Bot cost?

No standalone price has been announced. Grok Bot (SpaceXAI, announced Aug 11, 2026) is in beta and included with every SuperGrok plan (from $30/mo), Cursor Pro ($20/mo) and all Cursor Teams plans (from $40/seat/mo) as of Aug 26, 2026 — extended from SuperGrok Heavy, Cursor Ultra and Cursor Teams Premium at launch — on desktop (macOS) and iOS, with enterprise activation in the admin dashboard since Sept 3, 2026 (sources: x.ai/news/grok-bot-more-plans, Aug 26, 2026; cursor.com/pricing, checked Sept 20, 2026). It runs always-on agents on a cloud computer, so jobs keep going when the laptop is closed, and no usage caps have been published. Because there is no per-token meter, this calculator treats it as a bundled line item rather than a usage-modeled cost — budget the eligible subscription, not a token bill. Official announcement →

How much can model choice change AI coding agent cost?

Dramatically. On credit plans, premium models can burn credits up to 8× faster than standard ones — a task that costs one credit unit on a standard model might cost eight on a frontier model (TrueFoundry, Aug 6, 2026). Annual billing typically discounts 15–20%, and free/entry tiers often carry quotas teams exceed within weeks. That's exactly why this calculator exposes Model Strategy (open-weight vs. hybrid vs. frontier) and Delivery Risk as separate levers: the model picker is a cost lever, not just a settings control.

What is AI supplier risk and should it change my cost estimates?

AI supplier risk is the chance that a model provider's release timeline, safety posture, or roadmap shifts in ways that affect your stack's cost or availability. Real-world example: on Aug 18, 2026, OpenAI announced a two-week pause in reinforcement-learning training on its latest deployment-bound models and held its largest planned frontier RL run on hold, after internal evaluations flagged its upcoming Astra model at the "Critical" cyber-capability threshold under its Preparedness Framework — core Astra training continued and no model was cancelled. No published API prices changed, so this calculator's math stands, but provider roadmaps can shift and the ~20% monitoring overhead on frontier inference may add modest price pressure over time. Guidance: when evaluating AI costs and risks, include fallback AI models in your stack and review each provider's tooling roadmap before locking in long-term pricing or client deliverables. (Dated Aug 19, 2026; source: OpenAI — Pacing model development in an era of cyber-critical capabilities.)

What is the OpenAI safety-monitoring overhead toggle and should I turn it on?

The toggle is a stress-test switch (OFF by default) in the Model Strategy section. When ON, it applies a +20% multiplier to OpenAI-based strategies — GPT-5.6 Sol and Paid frontier only (Claude, GPT) — to model what would happen if OpenAI's safety-monitoring compute overhead were priced into API costs. The basis: on Aug 18, 2026, OpenAI estimated its expanded chain-of-thought safety monitoring adds roughly 20% overhead to the inference compute it monitors — required for all RL training and evaluations involving tools for models of Sol capability or higher, plus all Astra inference with tools after its Aug 7 Critical cyber-capability determination. Two things keep this a what-if, not a price change: OpenAI is not currently billing customers for the overhead (spokesperson via The Register, Aug 19, 2026; corroborated by TNW), and Anthropic says its safeguards make a similar slowdown unnecessary (Axios, Aug 19, 2026). The toggle has no effect on Anthropic (Claude), Google (Gemini), DeepSeek, open-weight, or local strategies — use it to stress-test an OpenAI-heavy quote, then leave it OFF for today's pricing. (Source: OpenAI — Pacing model development in an era of cyber-critical capabilities, Aug 18, 2026; verified via research brief t_b34197c5.)

How should I price ChatGPT Ads management for clients in the EU?

Price it as a pilot → retainer with a management fee on ad spend. OpenAI expands ChatGPT Ads to 31 European markets on Aug 24, 2026 (40 markets total); ads appear on Free and Go tiers only (Go ≈ €8/mo; Plus/Pro/Enterprise ad-free), and buying is agency-led first — advertisers go through the OpenAI Ads Solutions team, agency partners, and tech partners, with a self-serve Ads Manager following "later this summer." A defensible package: $1,500–$5,000 setup (account structure, conversion tracking/oCPC, GDPR-compliant targeting, creative prep) + 15–25% of ad spend as monthly management fee, floored at $2,000/mo to cover the manual review, policy monitoring (rules changed 4× in 5 months), creative iteration, and reporting — and capped around $7,500/mo as a typical rollout retainer. Ad spend is pass-through: the client pays OpenAI directly, so the management fee is the agency's revenue. EU campaigns cannot rely on Google/Meta-style personalization (GDPR consent requirements), so lead with conversation context and brand safety; health, mental-health, and political content are ineligible; under-18 excluded. Measurement is early — run a 2–3 month instrumented pilot against a Google AI Mode baseline before promising scale. Source: OpenAI — ChatGPT Ads expands across Europe (Aug 18, 2026) · Search Engine Land · Dataconomy — verified via research brief t_ff98639f.

How much will Meta's Hatch AI agent cost?

Reported, not final. The Information (Jun 4 + Aug 24, 2026, via PYMNTS and RuntimeWire) reports Meta has considered a premium "Hatch Plus" tier priced up to $199.99/month (5–10× the daily capacity of a free tier), but final pricing is NOT set and Hatch has not shipped as of Aug 25, 2026. Launch is targeted for late August or early September 2026, with a new model codenamed Watermelon targeted for October. Treat $199.99 as a reported/rumored anchor, not official pricing, until Meta announces. See the Meta Hatch AI Agent Price explainer for the full breakdown and the consumer comparison table above.

What does Ramp's business-spend data show about Claude Fable 5, Opus 5, and GPT-5.6 Sol adoption?

Ramp's August 2026 AI Index (businesses on its AI cost tracker, July 2026) shows flagship list price is not what most business workloads run on. Anthropic's flagship Claude Fable 5 — the July 2026 model, superseded by Fable 5.1 on Sept 1, 2026 at the same $10/$50 with $0.25 cache reads — drew only ~6% of Anthropic business tokens and 11.4% of model-attributed Anthropic spend: the most expensive model by far, but a minority of usage. OpenAI's GPT-5.6 Sol (currently $4/$20 per 1M promo through Nov 21, 2026) holds ~25% of OpenAI tokens and 23% of OpenAI spend, at roughly half of Fable 5's price. Anthropic's cheaper Claude Opus 5 ($5/$25 per 1M, exactly half of Fable 5, launched late July 2026) has already overtaken Fable 5 in enterprise spending, per FT-cited Ramp data (via Superpower Daily / CIO) — note Ramp's own AI Index does not state the overtake, and WinBuzzer cautions Opus arrived too late in the measured month to explain the full July pattern. Agency takeaway: price retainers against a realistic model mix (Opus 5 / Sol / Sonnet 5 capture the volume), not against flagship list prices. Select Claude Opus 5 in the Model Strategy dropdown to model this stack. Sources: Ramp AI Index Aug 2026 · WinBuzzer · BreezyScroll · Superpower Daily (FT-cited) — verified via research brief t_06f1cfb2.

What does the AgentCore Migration Estimator calculate?

The AgentCore Migration Estimator models the monthly platform-bill delta between Bedrock Agents Classic and AgentCore — runtime ($0.0895/vCPU-hr + $0.00945/GB-hr active), gateway ($0.005/1K invocations), memory ($0.25/1K short-term events, $0.75/1K records/month long-term), web search ($7.00/1K queries), policy ($0.000025/request), plus the shared model-inference tokens — and adds the one-time migration labor range by agent complexity (hours–1 day simple, 1–3 weeks custom orchestration, 3–6 weeks multi-agent). Classic itself is free as a platform; the delta is typically cents while labor is the number that moves the quote. No EOL, no deadline — the migration is a strategy choice, not an emergency. Full analysis: Bedrock Agents Classic → AgentCore Migration Cost.

What is GPT-6 Astra and how is it priced?

GPT-6 Astra is OpenAI's flagship model launched Sept 3, 2026 (the first model to reach Critical tier under OpenAI's Preparedness Framework, built for complex reasoning, coding, computer use, research, and document creation). Standard API pricing: $10 per 1M input / $50 per 1M output / $1 per 1M cached input ($12.50 cache writes). The 272K rule: prompts with more than 272K input tokens reprice the FULL request at 2x input and cache rates and 1.5x output — $20 input / $75 output / $2 cached / $25 cache writes per 1M. Context window: 1,050,000 tokens (922K max input, 128K max output); Batch/Flex are 50% of standard, Fast mode 2x. Availability (Sept 4): live in the OpenAI API and in ChatGPT Work and Codex for all Pro, Enterprise, and Business Premium users per OpenAI's developer account; showing up in ChatGPT as GPT-6 Pro for Pro and Business first, with Plus chat access to follow. Astra Pro — the higher Work/Codex tier — is available to Pro, Business, and Enterprise plans; Enterprise admins enable it for their workspace (off by default at launch). Select GPT-6 Astra in the Model Strategy dropdown to model this stack, and use the GPT-6 Astra context scenario ↓ to see sub-272K vs above-272K cost math. Full analysis: GPT-6 Astra API Pricing: $10/$50 per 1M, 272K Doubling, 1.05M Context.

What percentage of revenue does AI inference cost?

For the one operator who has published the number, AI inference runs at about 6.4% of revenue. On Sept 9, 2026 @levelsio reported that his profit margin is 99.4% without AI inference and 93% with it (mostly Photo AI) — a 6.4-point gross-margin cost, and roughly 91% of his total cost of goods. It is one self-reported, unaudited data point, so treat it as an anchor rather than a benchmark, and run your own numbers in the inference-as-a-margin-line estimator above.

How much do parallel agent teams cost an agency?

Two lines, and only one of them is published. The seat licence is the visible line: Cursor Teams is $40 per user/month (Standard) or $120 per user/month (Premium, 5x Agent limits), with individual plans at $20 / $60 / $200. The second line is what the agents themselves burn: Cursor bills cloud agents at API pricing for the selected model with a required spend limit (over-limit usage on-demand at the same rates) — and it publishes no price for Cursor Projects (beta, Sept 10, 2026), no compute-minute metric and no credits currency, so there is no vendor rate to multiply by an agent count. Model it as concurrent cloud agents × token spend per agent: the Parallel Agent Compute Cost Estimator on this page keeps the seat line and the compute line separate, and on its default inputs (5 seats × $40, 4 agents × $60) the $200 seat line is under half of a $440/month bill — an effective $88 per seat, 2.20× the sticker price. The only published anchor for the compute side is Cursor's individual-plan usage guidance: $60–$100/month for daily agent users, $200+/month for multiple agents or automation. Treat that as guidance, measure your own token spend, and never quote a fixed retainer off the seat price alone. Full breakdown: AI coding tool costs — seat vs compute.

AI Agency Pricing Calculator — Instantly Calculate Your Rate

AI Pricing Calculator

Service Type

Client Business Size

Number of Workflows / Automations

Your Experience Level

Project Timeline

Model Strategy (API Cost Profile)

Model pricing notes & optional adjustments

OpenAI safety-monitoring overhead (~20%) — stress-test what happens if OpenAI's estimated ~20% overhead on monitored inference compute (Sol-class tool-enabled RL training/eval and Astra tool-use inference) were priced into API costs. Applies a +20% multiplier to OpenAI-based strategies (GPT-5.6 Sol, Paid frontier). OpenAI is not currently billing customers for this overhead (spokesperson via The Register, Aug 19, 2026), and Anthropic says its safeguards make a similar slowdown unnecessary — so this toggle is OFF by default and does not affect Anthropic / Google / DeepSeek / open-weight / local strategies.

Sources: OpenAI — "Pacing model development in an era of cyber-critical capabilities" (Aug 18, 2026) · The Register (Aug 19, 2026) · TNW (Aug 19, 2026) · Axios (Aug 19, 2026) · Anthropic RSP v3 — verified via research brief t_b34197c5.

NEW Sept 10, 2026: DeepSeek V4.1-Flash — API name deepseek-flash, 552B MoE with native vision and a 1M-token context, MIT weights: off-peak $0.15 input / $0.60 output / $0.003 cached input per 1M (peak is exactly 2×). That cached-input rate is 100× below the peak input rate and ~83× below Claude Fable 5.1's $0.25 cache read; the row, the spread and a worked example are rendered from the calculator's own price data in the V4.1-Flash section . The retired IDs deepseek-v4-flash and deepseek-v4-flash-vision-exp are aliases: still accepted, served by V4.1-Flash, billed at the Flash price. DeepSeek V4 Pro is NOT retired — the announced Sept 14 reroute was withdrawn on Sept 11 and V4 Pro keeps its own row at unchanged billing; do not publish a migration deadline. Verify api-docs.deepseek.com/quick_start/pricing before quoting. Full analysis: DeepSeek V4.1-Flash pricing: the 100x cached-input spread and the withdrawn Sept 14 V4 Pro reroute

NEW Sept 11, 2026: The three-layer AI bill: model + tools + orchestration, worked — the model rate is only one of the three lines an agency pays: tool calls, orchestration/planner tokens and the retry loops they cause are priced out line by line against a real agent workload. Full breakdown: The three-layer AI bill: model + tools + orchestration, worked

DeepSeek price increase LANDED (Aug 16, 2026, 16:00 UTC): DeepSeek's official peak/off-peak rates are now published — off-peak V4 Pro $0.66 cache miss / $1.98 output / $0.022 cache read per 1M tokens (peak: $1.32 / $3.96 / $0.044). The old V4-Flash $0.14/$0.28 and V4-Pro $0.435/$0.87 rates are superseded. DeepSeek still holds the lowest cache-read tier on this page, but the current floor is DeepSeek V4.1-Flash at $0.003/1M off-peak (deepseek-flash, Sept 10, 2026) with V4 Pro second at $0.022/1M off-peak — and the "276x cheaper" headline compared Fable 5's $1.00 to DeepSeek launch pricing — with Fable 5.1's cache reads now $0.25 the gap is ~11x off-peak. Verify before quoting: DeepSeek pricing page · @JulianGoldieSEO cache-gap post

NEW Aug 16, 2026: DeepSeek V4 Pro (cache-aware) — a model strategy built for long agent runs that re-read context on every step. At a 92% cache hit rate (an assumption for agent workloads, not a published OpenRouter stat), the effective input cost collapses: 92% of input tokens bill at the $0.022/1M cache-read rate instead of $0.66/1M. Select this to model a cache-heavy agent stack, then compare against Claude Fable 5.1 ($10/$50 per 1M, $0.25 cache read) to see the gap — see the cache-aware scenario .

NEW Sept 1, 2026: Claude Fable 5.1 (Anthropic) — Anthropic's premium flagship, GA Sept 1, 2026 (API id claude-fable-5-1): $10 per 1M input / $50 per 1M output, $0.25 per 1M cache read (down 75% from Fable 5's $1.00; 0.025x base input). 1M-token context, 128K max output. Claude Mythos 5.1 is the same underlying model with cyber/bio-tuned safeguards — trusted access only, not GA. This is the most expensive per-token stack on the page — model it as the premium comparison when you're quoting clients whose workloads re-read large instruction sets on every agent step. Ramp AI Index (July 2026, measured the predecessor Fable 5): Fable 5 drew only ~6% of Anthropic business tokens and 11.4% of model-attributed Anthropic spend — flagship price, minority usage. Anthropic announcement · Claude pricing docs

NEW Aug 26, 2026: Claude Opus 5 (Anthropic) — launched late July 2026 at $5 per 1M input / $25 per 1M output (exactly half of Fable 5.1), delivering near-Fable-5.1 performance at half the price. Per FT-cited Ramp data (Aug 2026), Opus 5 has already overtaken Fable 5 (the predecessor) in enterprise spend after its late-July debut. Select this to model Anthropic's value-tier frontier stack — the one business token volume is moving to. Anthropic Opus

Grok 4.6 (SpaceXAI, released Aug 12, 2026): $2/$6 per 1M tokens (input/output), 500k context (Artificial Analysis), Intelligence Index 61 — matching GPT-5.6 Sol (max) at 50% lower input price and 70% lower output price vs Sol's current $4/$20 (promo window Aug 21 – Nov 21, 2026; previously $5/$30). A fast variant costs 2x; cache hits $0.50 per 1M input (-75%). Available in Cursor, Grok Build, the SpaceXAI API, OpenRouter, Vercel, and Cloudflare. x.ai announcement · Artificial Analysis · Agency impact analysis

NEW Sept 2, 2026: Gemini 3.8 Flash (Google, released Sept 2, 2026) — Google's best reasoning & coding model yet, the third Flash-family release in six weeks (3.6 3.7 3.8) and its recommended model for software engineering, autonomous agents, and complex multi-step reasoning. Intro pricing $0.75/$3.75 per 1M tokens (input/output, output includes thinking tokens) — valid through 2026-12-31, then $1.50/$7.50 from Jan 1, 2027 (identical structure to the 3.7 Flash intro it extends). Same 1M-token context and 64K max output; available to AI Pro/Ultra subscribers in the Gemini app, AI Mode, and Sheets, and to developers via Gemini API / AI Studio / Antigravity. Google positions it above 3.7 Flash on DeepSWE v1.1 (long-horizon software engineering) and in agentic benchmarks, noting it can use more tokens at higher effort levels. A separate Gemini 3.8 Flash Cyber variant (limited access via the Fairwind program) is not generally available — do not price it into client stacks. Google announcement · Google blog · Model card · Ars Technica · 9to5Google · Gemini 3.8 Flash API pricing post

LIVE Sept 4, 2026 (availability refresh): GPT-6 Astra (OpenAI's new flagship, announced Sept 3, 2026) — the most expensive OpenAI per-token stack on this page: $10 per 1M input / $50 per 1M output, $1 per 1M cached input, $12.50 cache writes. Its 272K rule reprices any prompt over 272K input tokens at 2x input and cache rates and 1.5x output for the FULL request ($20 input / $75 output / $2 cached / $25 cache writes per 1M), with a 1,050,000-token context window (922K max input, 128K max output). Availability (Sept 4): live in the OpenAI API and in ChatGPT Work/Codex for Pro, Enterprise, and Business Premium (OpenAI developer account confirmation); showing up in ChatGPT as GPT-6 Pro for Pro/Business plans first, Plus chat to follow; Enterprise admins enable Astra for their workspace (off by default at launch). Astra Pro — the higher Work/Codex tier, priced separately in Work/Codex credits — is available to Pro, Business, and Enterprise plans. Select this strategy to model the premium OpenAI flagship, and see the GPT-6 Astra context scenario for sub-272K vs above-272K cost math. Full pricing analysis

PREVIOUS VERSION (Aug 13 – Sept 1, 2026): Gemini 3.7 Flash (Google, released Aug 13, 2026) — Google's most intelligent workhorse model yet for coding and agents at its release, based on Gemini 3.6 Flash (released ~3 weeks earlier). Intro pricing $0.75/$3.75 per 1M tokens (input/output, output price includes thinking tokens) — valid through 2026-12-31, then reverts to $1.50/$7.50 from Jan 1, 2027 (its post-intro price equals 3.6 Flash's launch price; intro is exactly half). 1M-token context, 64K max output, 5,000 free search requests/mo (shared across Gemini 3.x), 50% batch discount, and a free tier for input/output. vs Gemini 3.6 Flash: FrontierCode 1.1 Main 43.6% (vs 34.4%), DeepSWE v1.1 65.3% (vs 49.0%), WebDev Arena Elo 1588 (vs 1538). Google blog · Gemini API pricing · Model card · Agency impact analysis

NEW Aug 13, 2026: Claude Sonnet 5 (Anthropic) — $2 per 1M input tokens / $10 per 1M output tokens, PERMANENT as of Aug 10, 2026. Anthropic made the launch pricing permanent and cancelled the previously scheduled Sept 1, 2026 increase to $3/$15 — no scheduled price change is pending (verified Aug 10, 2026 announcement + official pricing docs; see the FAQ below). Cache: $0.20 per 1M cached input; cache write $2.50 (5m) / $4 (1h) per 1M. Input at open-weight parity (Qwen 3.8 Max $2), output well below premium frontier (GPT-5.6 Sol $4/$20 promo) — a hosted frontier API with no open-weight/self-host upside. Anthropic announcement · Claude pricing docs

NEW Aug 22, 2026: GPT-5.6 Sol (OpenAI frontier, powers both Instant and deep reasoning for ChatGPT Plus/Pro with a reasoning-effort slider) — official API pricing published Aug 21, 2026: $4 per 1M input tokens / $20 per 1M output tokens (cached input $0.40 per 1M), a promotional rate valid Aug 21 – Nov 21, 2026 (down from $5/$30; API + eligible ChatGPT Work/Codex credits; ChatGPT Pro/Plus/Business subscription pricing unchanged). This option uses the verified promo rate. NEW Aug 26, 2026 (Ramp AI Index): Sol holds ~25% of OpenAI business tokens and 23% of OpenAI spend (July 2026) — the OpenAI workhorse for business workloads, at roughly half of Fable 5.1's price. See OpenAI pricing

Open-weight option: Alibaba's Qwen 3.8 Max (GA Aug 2–3, 2026; 2.4T-param MoE, ~95B active; 1M context) lists at $2/$6 per 1M tokens — open weights landed on Hugging Face Aug 12, 2026, so self-hosted cost now varies by hardware/quantization. License terms finalized: Qwen 3.8 Max's official license requires a separate Qwen license before commercial use only if you run a Model-as-a-Service or AI Work Assistant business AND aggregate revenue exceeds US$50M over any consecutive 12 months (no numeric % — individually negotiated; internal use exempt). Do not model it as free/unrestricted at scale. Kimi K3 ($3/$15) and GLM-5.2 (MIT) round out the open-weight class. See the open-weight section

NEW Aug 26, 2026: GLM-5.3-Flash (Z.ai) — the anonymous Ox Alpha stealth model (free on OpenRouter/OpenCode from Aug 20) was confirmed Aug 26, 2026 as GLM-5.3-Flash: 320B total / 18B active MoE, natively multimodal (text/image/video), 1M-token context, MIT-licensed weights on Hugging Face, served entirely on Chinese AI chips during the preview week. Official API pricing: $0.15 per 1M input / $0.50 per 1M output (cached input $0.03), with a 50% launch promo to $0.075/$0.25 through Sep 9, 2026 — roughly one-tenth of the flagship GLM-5.3 rate ($1.40/$4.40) and below DeepSeek V4 Pro off-peak on both axes. Benchmarks (DeepSWE v1.1 63.4, Terminal-Bench 2.1 84.3, AutomationBench 48.8, Z.ai Code Bench 29.0 vs Claude Opus 4.8's 29.5) are vendor-reported — no independent replication yet. MIT license means no Qwen-style licensing threshold, but self-hosting 320B/18B is a serious hardware project (SGLang/vLLM/TokenSpeed/KTransformers), not a consumer-GPU pull. Full analysis: GLM-5.3-Flash Pricing: Ox Alpha Revealed at $0.15/M

NEW Aug 28, 2026: Tencent Hy4 preview — Tencent's largest open-weight release yet: 770B total / 49B active MoE with a native 1M-token context (64K max output), Apache 2.0 weights on Hugging Face (tencent/Hy4-preview + FP8 variant). OpenRouter lists it at $0.834 per 1M input / $2.501 per 1M output (cached input $0.042 — a ~20x cache discount for agent workloads), with Tencent Cloud the sole provider. That is roughly 40% below GLM-5.3 on input and 43% on output, and ~6x below Kimi K3 on output — still pricier than DeepSeek V4 Pro off-peak ($0.66/$1.98). Tencent's internal blind eval (163 experts, 203 tasks) scored it 2.99/4 vs GLM-5.3 2.92 and Kimi K3 2.94 — tier parity, not a decisive win. Caveats: it is a preview (Tencent flags over-reasoning and over-verification), observed throughput is ~36 tok/s P50, first-3-day OpenRouter availability was ~86%, and the FP8 weights are ~770GB — not self-hostable on typical infra (8xH200-class or multi-node). Full analysis: Tencent Hy4 Preview Pricing: 770B Open Weights at $0.834/M

NEW Aug 28, 2026: Qwen3.8-Flash — Qwen's new open-weight MoE (open-sourced Aug 26, 2026 on Hugging Face + ModelScope; production API on QwenCloud same day): 125B total / ~6B active per token, multimodal, ~95% sparsity, 262,144-token native context 1M via YaRN, qwen-community-1.0 license (not Apache-2.0). Official pricing: $0.15 per 1M input / $0.47 per 1M output (cache read $0.016) on QwenCloud, live on OpenRouter at the same rate — below GLM-5.3-Flash at list on output and cache, below DeepSeek V4 Pro off-peak on both axes, and one of the lowest cache-read rates on this page. Artificial Analysis medians: $0.30/$1.20. Caveats: benchmarks are vendor-reported with no independent replication yet, this is a "Next" preview of the Qwen4 architecture (expect rapid iteration), and the license is a new qwen-community-1.0 layer, not MIT/Apache. Full analysis: Qwen3.8-Flash-Next Cost: Open-Weight MoE at $0.15/M

PRICING-PATH ALERT Aug 28, 2026: OpenRouter removed all 35 openai/*:batch model slugs on Aug 26, 2026 — completions now return 404 "No endpoints found" even though the slugs still answer metadata requests. The base models are untouched (openai/gpt-5 still routes at $1.25/$10.00 per 1M), so a dead :batch slug silently falls back to standard pricing at 2x the token cost unless you act. Three routes keep the 50% batch discount: OpenRouter's beta Batch API (plain slug, 24h window, text-only), the openai/flex tier (provider: {"only": ["openai/flex"]}, synchronous, GPT-5.6 family only — not gpt-5/4o/o3), or OpenAI's direct Batch API (still 50% off, 50K requests/batch, 24h completion). Audit every model string in your config with the endpoints check before quoting. Full routing guide: OpenRouter OpenAI Batch Endpoints Down: 3 Ways to Keep the 50% Discount

Local / self-hosted option: Meta Muse Glimmer (released Aug 10, 2026; 30B dense, Apache 2.0) runs a 4-bit quant (under 20 GB, official K-Quant-17GB GGUF = 16.76 GB) on a single consumer GPU (24 GB VRAM class) at >200 tok/s on an RTX 5090 with DFlash speculative decoding — marginal inference cost collapses to electricity once hardware is amortized. Official local runtime: Ollama (ollama run muse-glimmer — official build 18 GB / 128K context / text+image; Apple Silicon MLX build muse-glimmer:30b-mlx, 21 GB). No official Meta hosted API price has been published; the same model is available hosted on Together AI at $0.35 / $1.50 per 1M tokens (input/output, $0.04 cached) — see the Local vs API cost scenario . It is not frontier-grade on the hardest reasoning, the 17 GB quant carries ~1.0% benchmark degradation, and you own uptime/ops/security.

Delivery Risk (Failure / Retry)

Portability / Agent Plugins

Agent Plugins 1.0.0 (Aug 6, 2026) is the open standard for packaging Agent Skills + MCP servers into one portable plugin that runs across Codex, ChatGPT, Cursor, GitHub Copilot, Kiro, and VS Code. Build once, then apply a portability discount for each additional compatible client instead of charging full per-platform adaptation — and price plugin packaging & distribution as its own deliverable. How this changes pricing

Monthly Hours Saved for Client 40 hrs/mo

5 hrs150 hrs300 hrs

AI Agency Pricing Reference Table 2026

Service Type	Setup Fee Range	Monthly Retainer	Avg Margin	Best For
Chatbot / Assistant	$1,500–$5,000	$500–$1,500/mo	65–75%	SMBs, e-commerce, service cos
Email Automation	$2,000–$6,000	$750–$2,000/mo	60–72%	Coaches, SaaS, agencies
Lead Generation Bot	$3,000–$8,000	$1,000–$3,000/mo	55–70%	Real estate, insurance, finance
Content Automation	$2,500–$7,500	$800–$2,500/mo	65–80%	Content creators, media, blogs
Full Office Automation	$8,000–$35,000	$2,500–$7,500/mo	45–65%	Mid-market, growing teams
Custom AI Agent	$5,000–$25,000	$1,500–$5,000/mo	50–70%	Tech cos, SaaS, operations
Social Media Automation	$1,500–$4,500	$600–$1,800/mo	70–82%	Brands, coaches, ecommerce
ChatGPT Ads Management (EU)	$1,500–$5,000	15–25% of ad spend ($2K–$7.5K/mo floor)	55–70%	SMB advertisers targeting the EU rollout (Aug 24, 2026 — 31 markets); agency-led buying window

New Sept 3, 2026: GPT-6 Astra (OpenAI flagship) is the page's new most-expensive OpenAI tier — $10/$50 per 1M with $1 cached input, and any prompt over 272K input tokens reprices the full request at $20/$75 per 1M (1.05M context window). Model it in the calculator's Model Strategy dropdown — see GPT-6 Astra API pricing and the GPT-6 Astra context scenario . * Ranges reflect 2026 US market rates. Final pricing depends on complexity, client size, and your experience level. Model strategy affects margins more than list prices: open-weight stacks (Qwen 3.8 Max at $2/$6 per 1M tokens, Kimi K3, GLM-5.2) cut the compute line vs. paid frontier APIs — and Grok 4.6 (SpaceXAI, released Aug 12, 2026) now lists at the same $2/$6 per 1M tokens as Qwen 3.8 Max with a 500k context window and an Artificial Analysis Intelligence Index of 61 (matching GPT-5.6 Sol max) — 50% below GPT-5.6 Sol's current $4/$20 per 1M (promo window Aug 21 – Nov 21, 2026; previously $5/$30) on input tokens and 70% below on output tokens (x.ai · AA ). New Aug 28, 2026: Tencent Hy4 preview (770B MoE, 49B active, Apache 2.0, native 1M context) lists at $0.834/$2.501 per 1M tokens on OpenRouter — ~40% below GLM-5.3 on input, ~43% below on output, and ~6x below Kimi K3 on output, though still above DeepSeek V4 Pro off-peak; a preview with over-verification tendencies and ~770GB FP8 weights (not typical self-host) — model it in the calculator's Model Strategy dropdown (OpenRouter · HF · Full analysis ). New Aug 28, 2026: Qwen3.8-Flash (Qwen, open weights Aug 26, 2026; 125B MoE / ~6B active, ~95% sparsity, qwen-community-1.0 license) lists at $0.15/$0.47 per 1M tokens (cache read $0.016) on QwenCloud and OpenRouter — below GLM-5.3-Flash at list on output and cache, below DeepSeek V4 Pro off-peak on both axes, and one of the lowest cache-read rates on this page; Artificial Analysis medians $0.30/$1.20; benchmarks vendor-reported, 'Next' preview variant — model it in the calculator's Model Strategy dropdown (OpenRouter · QwenCloud · Full analysis ). New Aug 13, 2026: Claude Sonnet 5 (Anthropic) — $2/$10 per 1M tokens, PERMANENT (Aug 10, 2026); the previously scheduled Sept 1, 2026 $3/$15 increase is CANCELLED, no scheduled change pending — input at open-weight parity, output well below premium frontier (Anthropic · Pricing docs ). New Sept 2, 2026: Gemini 3.8 Flash (Google) is Google's best reasoning & coding model yet — the third Flash release in six weeks at the same intro pricing $0.75/$3.75 per 1M tokens (valid through 2026-12-31, then $1.50/$7.50; 1M context, 64K max output), recommended for software engineering, autonomous agents, and complex multi-step reasoning, and now the current workhorse Flash option on this page (3.7 Flash, which carried the same intro rate from Aug 13, is preserved above as a previous version) (input/output, valid through 2026-12-31, then $1.50/$7.50; 1M context, 64K max output) — undercutting Qwen 3.8 Max and Grok 4.6 on both axes during the intro window (Google blog · Gemini API pricing ). Qwen 3.8 Max license terms finalized (Aug 12, 2026): separate Qwen license required before commercial use only if you run a Model-as-a-Service or AI Work Assistant business AND aggregate revenue exceeds US$50M over any consecutive 12 months (no numeric % — individually negotiated; Reuters' "up to 30%" is reporting, not license text; internal use exempt) — threshold 2.5x higher than Moonshot's Kimi K3 ($20M/12mo), but with a new AI Work Assistant trigger category (coding/office-productivity agents like Qoder, QwenWork) that Kimi K3 lacks; display clause identical (100M MAU or $20M/month revenue prominently display the model name). DeepSeek price increase LANDED (Aug 16, 2026, 16:00 UTC) — official peak/off-peak rates are live: off-peak V4 Pro $0.66/$1.98 per 1M tokens with $0.022 cache reads (peak $1.32/$3.96/$0.044); pre-hike V4-Flash $0.14/$0.28 and V4-Pro $0.435/$0.87 are superseded. Even after the increase, V4 Pro cache reads are ~11x below Claude Fable 5.1's $0.25/1M (Fable 5.1 cut cache reads 75% from Fable 5's $1.00 on Sept 1, 2026) — still the dominant line for long agent runs (see the cache-aware scenario below; the ~276x figure from @JulianGoldieSEO compared Fable 5's $1.00 to DeepSeek launch pricing and is stale on both sides). Verify api-docs.deepseek.com/quick_start/pricing before quoting. New (Aug 11, 2026): the cheapest per-workload option is now local self-host — Meta's Muse Glimmer (30B dense, Apache 2.0, 4-bit quant under 20 GB on a single consumer GPU) runs at electricity-cost marginal inference once hardware is amortized, undercutting every hosted API at sustained volume (Local vs API cost scenario ). New (Aug 11, 2026): for Content Automation work with image deliverables, Microsoft's MAI-Image-2.6 (announced Aug 10, 2026) is now #2 on the Arena text-to-image leaderboard (mai-image-2.6-preview, Elo 1336 ±11) — ahead of Google, Meta, and xAI, behind only OpenAI's GPT-Image-2 — with text rendering the standout gain (+91 Elo vs 2.5) for packaging, signage, and ad-copy image work. Availability: Arena now, MAI Playground this week, Foundry API soon; Foundry pricing for 2.6 is not yet published, so treat prior-gen MAI-Image-2.5 (~$48/1k images) as the reference until rates drop. GPT-5.6 Sol (Aug 6, 2026) now powers both Instant and deep reasoning for ChatGPT Plus/Pro — one consistent model with a reasoning-effort slider; GPT-5.6 Luna is the new default for Free/Go users (unlimited text chats rolling out this week/next week). NEW Aug 22, 2026: OpenAI published official GPT-5.6 Sol API pricing Aug 21, 2026 — $4/$20 per 1M input/output tokens (cached input $0.40 per 1M), a promotional rate valid Aug 21 – Nov 21, 2026 (down from $5/$30; API + eligible ChatGPT Work/Codex credits; ChatGPT Pro/Plus/Business subscription pricing unchanged) — verify current rates at OpenAI pricing before quoting. Agent Plugins 1.0.0 (Aug 6, 2026) adds a portability lever: one plugin package serves Codex, ChatGPT, Cursor, GitHub Copilot, Kiro, and VS Code, so setup fees carry a build-once discount and plugin packaging & distribution is priced as its own deliverable (see the Portability selector and Agent Plugins section ). New (Aug 2026): agent payment rails like Cloudflare Wallets (announced Aug 4, 2026) add a creator-set spending cap at the wallet layer — the same rails that make the capped vs uncapped spend comparison in this calculator worth running before you quote a retainer. New Aug 26, 2026: Claude Opus 5 (Anthropic) added as a selectable model strategy — $5 per 1M input / $25 per 1M output, exactly half of Fable 5.1 (launched late July 2026). Per Ramp AI Index (Aug 12, 2026), Fable 5 drew ~6% of Anthropic business tokens / 11.4% of Anthropic spend while GPT-5.6 Sol holds ~25% of OpenAI tokens / 23% of OpenAI spend at roughly half of Fable 5.1's price; FT-cited Ramp data (via Superpower Daily/CIO) reports Opus 5 overtook Fable 5 (the predecessor) in enterprise spend after its late-July debut (attributed to FT-relayed Ramp data — Ramp's own report does not state the overtake; WinBuzzer cautions Opus arrived too late to explain the full July pattern). Select Claude Opus 5 in the Model Strategy dropdown to model the value-tier Anthropic stack (Ramp AI Index · Anthropic Opus ).

AI Inference as a Margin Line

Inference as a Share of Gross Margin Consumed

On Sept 9, 2026, @levelsio (Pieter Levels) reported that without AI inference his profit margin is 99.4%, and that with AI inference — mostly Photo AI — it goes down to 93%. One cost line moved his gross margin 6.4 points: inference works out to ~6.4% of his revenue and ~91% of every dollar he spends to deliver the product. Read that way, inference stops being a “tool cost” you pass through to a client and becomes a share of the gross margin you keep. This estimator runs that conversion on your own numbers, with his case preloaded. The full step-by-step arithmetic, the honest caveats and the sources behind the reference case are in When AI inference is your biggest cost line.

Reference Case

Preloaded as his 6.4-point drop entered per $100,000 of monthly revenue. He has never disclosed revenue, and the drop is a ratio — any revenue base reproduces 99.4% 93.0% exactly. Edit any field to switch to Custom.

Monthly Revenue ($)

Monthly Cost of Goods EXCLUDING AI Inference ($)

Servers, hosting, third-party APIs, support — everything you pay to deliver the product except AI inference. The preset's $600 is the 0.6%-of-revenue residual implied by the 99.4% figure.

Monthly AI Inference Spend ($)

Model spend — tokens, images, audio, video, embeddings, fine-tunes. Include it here even if you resell it to clients at cost: this measures the margin it consumes before any markup.

Gross Margin WITHOUT Inference

99.4%

$99,400 of gross margin on $100,000 of revenue

Gross Margin WITH Inference

93.0%

$93,000 of gross margin on $100,000 of revenue

Margin Cost of AI Inference

6.4 pts

99.4% 93.0% gross margin

Inference as % of Gross Margin Consumed

6.4%

$6,400 of $99,400 gross margin

Inference as % of Revenue

6.4%

$6,400 of $100,000 revenue

Inference Share of All COGS

How to read this

Gross margin without inference: ($100,000 − $600) ÷ $100,000 = 99.4%. Adding $6,400 of inference takes cost of goods to $7,000 (7.0% of revenue), so gross margin falls to 93.0% — a 6.4-point margin cost. Measured against the margin you keep, that is 6.4% of gross margin consumed. Inference is 6.4% of revenue and 91.4% of all cost of goods — 10.7× the non-inference cost of goods.

Model: gross margin = (revenue − cost of goods) ÷ revenue, computed before and after adding monthly AI inference spend to cost of goods; “inference as % of gross margin consumed” = inference spend ÷ gross margin before inference. Reference case: @levelsio (Pieter Levels), Sept 9, 2026, 15:21:24 UTC — “Without AI inference my profit margin is now 99.4%… With AI inference (mostly Photo AI) it goes down to 93%!” (@levelsio on X ). Entered here per $100,000 of monthly revenue because he never disclosed revenue and the 6.4-point drop is a ratio that any revenue base reproduces. Caveats: one self-reported, unaudited data point from a single operator with one dominant image-generation workload (Photo AI); he said “profit margin”, not GAAP gross margin — a 0.6% residual cannot contain ~2.9% card-processing fees — and he gave no revenue base and no time period. Treat 6.4 points as an anchor from one business, not an industry benchmark.

Token Waste & Context Governance

Why the bill rises while the price falls

GPT-5.6 Sol is $4 per 1M input / $20 per 1M output — a 20% cut on input and a 33% cut on output, promotional at least through November 21, 2026. Prompts above 272K input tokens bill at 2× input and 1.5× output for the full request. Meanwhile agents re-send context on every call and retry loops burn tokens that produce nothing. This estimator prices the bill instead of the token: same tasks, two governance regimes. The arithmetic and its sources: why AI bills rise when prices fall.

Governance Scenario

Agent Tasks Completed per Month

Context Tokens per Model Call

System prompt + tool instructions + retrieved context. The work is held constant — this is the business workload.

Output Tokens per Call

Calls per Task (Context Re-Read Multiplier)

How many times one task re-sends its context. Ungoverned agents loop; 3× is the common blow-up. No vendor publishes an industry average, so the default is illustrative.

Wasted-Token Rate (% of Paid Tokens with No Accepted Output)

At 25% you pay 1.33× for the same accepted work. Illustrative default.

Context pruning / compaction on

Apply OpenAI's long-context pricing above 272K input tokens

Input Price ($ / 1M tokens)

Output Price ($ / 1M tokens)

Share of Re-Reads Served from Cache (%)

Optional. Cached input bills at $0.40/1M (10%) with writes at $5.00/1M (1.25×) — OpenAI's published rates. Needs a stable prefix.

Agents Running This Workload

Monthly Spend Cap per Agent ($, 0 = uncapped)

Governance-Adjusted Monthly Token Cost

$105,600.00

per completed task: $10.56 · 13.89× the per-token estimate

Per-Token Estimate (one clean pass)

$7,600.00

10,000 tasks × 150K context + 8K output at $4/$20

Governance Premium

$98,000.00

what governance would recover

Re-Read Premium

$64,800.00

7,500.0M input tokens re-sent

Wasted Spend

$26,400.00

3,000.0M tokens, 25% of what you pay for

Long-Context Cliff (>272K)

$51,200.00

every billed token at 2× input / 1.5× output

Billed Tokens / Month

12000.0M

320.0M output

Governance Multiple

13.89×

vs the per-token estimate

Cost per +1pt of Token Waste

$1,408.00

one more call per task: $59,280.00

Promo-Expiry Exposure (Nov 22, 2026)

Monthly Budget Exhausted

day 1.4

5 agents × $1,000 cap vs $105,600/month

Governed, pre-cut ($5/$30)

$12,505.26

same workload, same behaviour, list sheet

Price Effect (the Sol cut)

-$2,905.26

at governed volume — negative means the cut saved money

Volume + Governance Effect

+$96,000.00

the behaviour change at these inputs

Net Month-over-Month Change

+$93,094.74

price effect + volume effect, exactly

How to read this

Sources & assumptions. Prices, the 20%/33% cut, the 272K long-context rule, the 1.25× cache-write rate and the Nov 21, 2026 promo expiry are OpenAI's own model and pricing pages, read 2026-09-11. The $5.00/$30.00 list rate is from AWS's Bedrock price-reduction notice (2026-08-21). Re-read depth and waste rate have no published industry average — those defaults are illustrative; the prices and the cliff rule are not. This covers API consumption: if your team uses Sol through ChatGPT seats, the cut does not lower your bill. Fault, context and agentic-loop evidence: Bain (2026-09-10), IBM (2026-09-10), Microsoft Azure (2026-09-10), Unite.AI (2026-09-11).

Agent Failure & Retry Cost Estimator

What a Failed Loop Really Costs

Agent workflows rarely run clean the first time. On Aug 5, 2026, levelsio (Pieter Levels) reported burning $500 per Gauntlet Loop run — an AI-coding method that fans out subagents and loops until "utterly perfect" — then corrected it to $900 total with 95% of generated code removed. Measured baselines are ~$0.06 per request and "a few dollars per task"; failure modes (retry storms, subagent fan-out, silent misconfiguration) turn that into $500 loops and $2,000 overnight bills. Use this estimator to model what retries actually add to your spend.

Scenario Preset

The cap-breach preset models Cloudflare Wallets' manual-override rule: when an agent hits its creator-set allowance, over-limit requests are blocked and routed to a human for approval — that approval time is billable.

Base Cost per Loop Attempt ($)

Failed Loop Attempts (Retries)

Cleanup / Rework Cost ($)

Agent Wallet & Spend Cap Estimator

What Wallet Rails Really Cap

On Aug 4, 2026 Cloudflare announced Cloudflare Wallets and cloudflare.pay — the buyer side of agentic commerce. An Account Wallet (human-owned) funds Virtual Wallets (agent-owned) with guardrails: a spending cap / allowance, an approved merchant allow-list, and a maximum transaction size. Agents that hit a limit must request a manual override from a human — they cannot approve escalations themselves, and the cap is enforced at the wallet's API layer, not inside the model's system prompt, so prompt injection can't lift it. Cloudflare's own example: give every employee a $100/week budget for AI inference by provisioning one Account Wallet and Virtual Wallets per employee with that rule. Fees are undisclosed so far — this estimator defaults to 0% and lets you set your own. For agencies, the takeaway is cap-constrained planning: with wallet-layer limits, worst-case spend per agent becomes a known number (the cap), so cost forecasts get a hard ceiling instead of a tail.

Number of Agents

Monthly AI Spend per Agent (uncapped, $)

Wallet Fee (% of spend) — ESTIMATE

Cloudflare has not published wallet fees; this default is an estimate. Default 0% = no wallet fee until a fee schedule ships. Credit-card rails run ~1.5–3.5% + fixed fee; stablecoin/x402 micropayments can be sub-cent per transaction (~$0.0001) — set your own rail's fee here. Update this default when Cloudflare publishes fees.

Creator-Set Cap per Agent / Month ($)

Optional. Leave blank for uncapped. This is the Virtual Wallet allowance an Account Wallet owner defines (e.g. $100/week = ~$433/month per agent).

ChatGPT Business Seat Cost Estimator

Standard vs Premium Seats

On Aug 25, 2026 OpenAI launched Premium seats for ChatGPT Business (announced August 2026) — a higher-usage seat tier for teams that outgrow Standard. Premium seats cost $125 per user per month billed monthly, or $100 per user per month billed annually (a 20% annual discount); Standard Business seats remain $25 per user/month, or $20 billed annually. Premium includes 5x more usage than Standard, is not subject to the five-hour usage limit, and gets predictable weekly usage resets. A workspace needs at least 2 paid seats (any mix of Standard + Premium) and, since Aug 24, 2026, caps at 200 paid seats per subscription; larger deployments move to ChatGPT Enterprise (sales-led). Model a client's seat line here, then add it to the retainer math above.

Seat Type

Premium = 5x more usage than Standard, no five-hour usage limit, weekly usage resets.

Number of Seats (Users)

2-seat minimum (any mix of Standard + Premium) · 200-seat cap per subscription. Workspaces can mix both seat types.

Billing Cadence

All seats in a workspace share one cadence; changing cadence takes effect at renewal.

Parallel Agent Compute Cost Estimator

Seat Licence + Parallel Agent Compute, Priced Separately

The coding-agent bill has two lines, and the pricing page only shows one. A per-seat licence buys a person access — Cursor Teams is $40 per user/month (Standard) or $120 per user/month (Premium, 5x Agent limits). The coordinator + parallel-subagent pattern adds a second, usage-metered line on top of it: Cursor Projects (beta, launched September 10, 2026) puts a coordinator agent over the subagents it creates and manages — "the coordinator agent in a project doesn't write code itself; it plans the work, delegates it to agents that implement it, and brings the finished work back to you to check" — and those agents "run as many in parallel as the work needs", on the Project's own cloud computer rather than your laptop. Cursor bills that layer as API-priced tokens for the selected model, with a required spend limit (over-limit usage bills on-demand at the same rates). It publishes no Projects price, no compute-minute metric and no credits currency, so the per-agent rate below is a user-editable placeholder rather than a vendor rate. Model the seat line and the compute line here before you quote a fixed fee — see the seat-vs-compute comparison and the model + tools + orchestration frame this extends.

Coding-agent seats (licences)

People who hold a seat. This is the line a vendor's pricing page shows — and the one a seat count can price.

Seat licence price ($ per seat / month)

Published anchors (Cursor, verified 2026-09-15): Teams Standard $40/user/mo, Teams Premium $120/user/mo (5x Agent limits); individual plans Pro $20, Pro+ $60, Ultra $200. Replace with your own plan's per-seat rate.

Concurrent cloud agents (parallel subagents)

How many agents the coordinator runs at once. Cursor publishes no concurrency ceiling — coordinators run "as many in parallel as the work needs", with no quota and no rate — so this is your own planning figure, not a vendor limit.

Token spend per active agent ($ / agent / month) — PLACEHOLDER

No vendor rate exists for this line. Cursor charges cloud agents as API-priced tokens for the selected model with a required spend limit, so this field is user-editable and defaulted to a labelled placeholder: the default $60 sits at the low end of Cursor's published individual-plan usage guidance — $60–$100 per month for daily agent users, $200+/month for multiple agents or automation — which is plan guidance, not a published compute price. What drives the real figure: how many agents run in parallel, how much context each one re-reads on every step (larger context windows raise token usage and cost), the retries a failed fan-out costs, and the model rate those tokens bill at. A seat price is an access fee — it is identical whether one agent runs an hour or forty run all week — so seat pricing alone understates a multi-agent workflow: the compute line, not the seat line, is the delivery cost, and the results below show them separately plus the effective cost per seat once compute is counted.

Claude Code Usage Limit Cost Impact Estimator

Weekly Claude Code Limits: Net −17% Since Sept 14, 2026

Anthropic's temporary +50% weekly Claude Code boost ended September 13, 2026 at 11:59 PM PT, and the permanent level that replaced it took effect September 14, 2026 for Pro, Max, Team, and seat-based Enterprise plans: weekly limits are now 25% higher than they were before the promotion [1][2]. Against the allowance users actually had during the promotion that is a net −16.7% (≈ −17%) cut in weekly capacity — the canonical index runs pre-promotion baseline 100 promotion 150 permanent 125 [1][3][4]. Anthropic published both numbers itself: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." [3]. Pick the index you measure against below, enter the share of that weekly allowance your agency uses, and see where the workload lands on the permanent 125 index plus what it costs to hold the promotion-era capacity. Weekly limits only — 5-hour session limits, plan pricing, and API billing are unchanged [1][6]. Anthropic does not publish fixed weekly message/token counts for any plan, so this estimator is index-based (125/150 = 0.833) [1][6][9].

Plan

The permanent sentence names all four plan groups. It does not restate the promotion's exclusions, so no claim is made here about Free or consumption-based Enterprise seats. Pro/Max share one weekly pool with Claude apps; Team gives individual member limits.

Weekly allowance you are measuring against

The index replaces hour counts — Anthropic publishes no per-plan weekly figure. Default is the permanent 125 allowance, because that is the one in force today. Pre-promotion baseline 100 promotion allowance 150 permanent allowance 125.

Your weekly Claude Code usage (% of the selected index basis)

100% = you are using that basis's full weekly allowance. Anthropic does not publish absolute counts for any plan, so work in percentages of the index.

Seat price ($/month)

What you pay per seat today (example: a Max 5x seat at $100/mo covered the 150 index units during the promotion; holding that same weekly capacity on the permanent 125 index runs ≈ $120/mo). Arithmetic illustration, not a published price.

Number of seats

Claude Code seats on the plan. For Team / seat-based Enterprise, limits are per individual member, not a shared org pool.

Compute Supply Scenario — Power-Constrained vs Baseline

What Power Constraints Do to Your 2026–2027 Token Bill

On Aug 10, 2026 Anthropic, Macquarie Asset Management, and Singapore's GIC announced Theseus Infrastructure — a platform to develop, operate, and lease purpose-built U.S. data centers to Anthropic under long-term agreements, with Anthropic as anchor tenant and Macquarie/GIC owning the platform and funding the majority of each project's equity. Anthropic separately pledged to pay 100% of grid-upgrade costs for interconnecting its data centers through its own electricity charges. No capital commitment, capacity, or lease terms were disclosed. Why this matters to your quotes: Gartner forecasts global data center electricity up 26% in 2026 to 565 TWh, passing 1,200 TWh by 2030 where "grid supply may be insufficient" — power, not chips, is becoming the binding constraint on AI supply. Capacity is already being locked in years ahead: up to 5GW of AWS Trainium (more than $100B over 10 years) and ~3.5GW of Google TPU via Broadcom from 2027. This scenario models the supply-side risk the rest of the calculator prices away: a power-constrained per-token outlook vs your baseline rate. All adjustments are ESTIMATES — projections, not published prices.

Monthly Token Volume (millions)

Baseline Cost per 1M Tokens ($)

Your current blended rate (input + output). For a frontier-mix agent stack, $2–$5 per 1M is a typical range.

Horizon

Presets update the two ESTIMATE assumptions below; adjust them to your own read.

Grid / Power Pass-Through (% add-on) — ESTIMATE

Share of power-cost and grid-upgrade cost increases passed into per-token pricing. Gartner forecasts +26% YoY power demand in 2026; 2027 preset assumes a tighter grid.

Locked-In Capacity Relief (% discount) — ESTIMATE

Scarcity premium already removed by committed capacity: up to 5GW AWS Trainium + ~3.5GW Google TPU (from 2027). 2027 preset assumes these bind.

AgentCore Migration Estimator — Bedrock Agents Classic AgentCore Cost

What the Classic AgentCore Migration Really Costs

On July 30, 2026, AWS stopped opening Bedrock Agents Classic to new customers: the model catalog is frozen, and accounts without Bedrock Agents usage in the prior 12 months get HTTP 403 AccessDeniedException on CreateAgent / InvokeInlineAgent — with no exception process [1]. There is no end-of-life date and no migration deadline [1]. Classic itself is free as a platform; AgentCore bills ~12 consumption components — runtime $0.0895/vCPU-hr + $0.00945/GB-hr (active use only), gateway $0.005/1K invocations, memory $0.25/1K short-term events / $0.75/1K records/mo, web search $7.00/1K queries, policy $0.000025/request [3]. The one first-hand migration report showed a simple agent's bill moving ~$4.87 ~$5.02/month (+$0.15) — tokens dominate [10]. This estimator models the platform delta AND the one-time migration labor — the number that actually moves your quote.

Agent Complexity

Drives the one-time labor range: hours–1 day (simple), 1–3 weeks (custom), 3–6 weeks (multi-agent) [9].

# Agents

Sessions / Day

Avg Active Runtime per Session — vCPU-hr

Runtime bills only for active use; I/O wait is free [3].

Avg Active Runtime per Session — GB-hr

Tool Calls per Run ( Gateway invocations)

MCP tool exposure meters at $0.005/1K invocations [3].

Memory Events per Run

Short-term memory $0.25/1K events [3].

Long-Term Memory Records Stored / Month

Long-term storage $0.75/1K records/month [3].

Web-Search Queries / Day

$7.00/1K queries — a new billable component under AgentCore [3].

Policy Auth Requests / Day

$0.000025/request — effectively noise at agency scale [3].

Model (per 1M tokens, input/output)

Inference is the shared meter in both stacks — same tokens, same model; AgentCore allows any provider [1].

Avg Input Tokens per Run (thousands)

Avg Output Tokens per Run (thousands)

Blended Agency Rate ($/hr, for labor)

$150–$350/hr is the typical AI engineering band (see the cost-of-hiring guide).

Gemini API Cost & Model Routing Savings

Route Cheaper, Spend Less

Google Cloud's new managed model routing (API Gateway, Public Preview since Aug 3, 2026) accepts your existing OpenAI-compatible chat requests, inspects the model name in each payload, and routes the call to a cheaper foundation model — with no client-side code changes. This estimator shows the potential token-cost savings from routing simple traffic to Gemini Flash-Lite instead of paying Flash/Pro rates for everything.

Monthly Input Tokens (millions)

Monthly Output Tokens (millions)

Current Model

% of traffic routed to Flash-Lite 80%

0% (no routing)50%100% (all to Flash-Lite)

Google Cloud Model Routing: Cutting Gemini API Costs

What changed

On August 3, 2026 Google Cloud added managed model routing to API Gateway (Public Preview). It accepts OpenAI-compatible chat requests, transcodes them in-flight, and dispatches them to Gemini, Anthropic Claude, or OpenAI models hosted in Vertex AI Model Garden. Google positions it as a managed replacement for self-hosted proxies like LiteLLM — no proxy server to host, scale, or maintain.

How the savings work

Routing is driven by the model name in each request payload. You define a router with a default model plus rules mapping client model strings to cheaper backends — unmatched traffic falls back to the default. Example: send all traffic to Flash, set the default to Flash-Lite, and route only complex/agentic requests to Flash. Google's own examples use google/gemini-3.5-flash-lite, google/gemini-2.5-pro, anthropic/claude-opus-4-7, and openai/gpt-oss-120b-maas. The current Gemini 3.8 Flash (Gemini API model ID gemini-3-8-flash, released Sept 2, 2026) is Google's best reasoning & coding Flash model and the workhorse Flash option for coding/agent traffic — at intro pricing $0.75/$3.75 per 1M tokens (valid through 2026-12-31, then $1.50/$7.50), the same rate as the previous Gemini 3.7 Flash (Gemini API model ID gemini-3-7-flash, released Aug 13, 2026) is the workhorse Flash option for coding/agent traffic — at intro pricing $0.75/$3.75 per 1M tokens (valid through 2026-12-31, then $1.50/$7.50), it is already cheaper than 3.5 Flash on both axes, so routing simple traffic from it to Flash-Lite still saves but the gap is smaller.

Illustrative scenarios (directional)

Using Google's published list prices: a content agency sending 50M input + 10M output tokens/mo to Flash at $165/mo could route 80% to Flash-Lite and drop to ~$65/mo — ≈ $100/mo (~61%) saved. A multi-tier client setup on 2.5 Pro at $212.50/mo with 70% budget-tier traffic could drop to ~$100.85/mo — ≈ $111.65/mo (~53%) saved. A 5% fallback-traffic leak onto Flash-Lite instead of Flash saves ~$16/mo on that slice alone. Token volumes and split percentages are assumptions; substitute your own usage.

Caveats before you build on it

Public Preview: text-only, name-based routing to MaaS models in Model Garden; request-side streaming, gRPC, WebSockets, Gemini Live, VPC-SC, and Private Service Connect unsupported.
One-way mode: you cannot retrofit routing onto an existing gateway or remove it — switching requires a new API config + gateway.
Single-host constraint: all models in one router must share the same hostname (global or one regional endpoint).
Pricing gap: no model-routing-specific fee was found in the reviewed sources; confirm your exact model versions and region before quoting a client.
No per-request observability yet: routing decisions aren't attributed per request in logs during preview.

Sources

2026-09-15: Added the Parallel Agent Compute Cost Estimator (#parallel-agent-compute — kanban t_5de39b46, fact base research brief t_6687dbf2) — the calculator now prices the two lines a coding-agent bill actually has, separately. Cursor Projects shipped in beta on Sept 10, 2026 with a coordinator agent over parallel subagents that "run as many in parallel as the work needs", and Cursor bills that agent layer as API-priced tokens for the selected model with a required spend limit — it publishes no Projects price, no compute-minute metric and no credits currency, so the new inputs are seats × seat licence price (published anchors: Teams Standard $40/user/mo, Teams Premium $120/user/mo) and concurrent cloud agents × token spend per agent per month, the second field left user-editable and defaulted to a labelled placeholder ($60, the low end of Cursor's published individual-plan usage band of $60–$100/mo for daily agent users, $200+/mo for multiple agents or automation — presented as plan guidance, not as a published price). Outputs: per-seat licence line, parallel agent compute line, total (seat + compute), compute share of the bill and effective cost per seat. Zero-valued inputs reproduce the seat-only total exactly and hide the compute line. New section #parallel-agent-compute + calculateParallelAgents(), FAQ item + FAQPage schema entry (4445 Q), WebApplication featureList, meta keywords, Umami event parallel_agents_estimated, and a new node tests/parallel-agent-compute-verify.mjs harness that extracts the shipped function out of index.html. Cross-links the live seat-vs-compute comparison at /ai-coding-tool-costs and extends (does not restate) the model + tools + orchestration frame at /orchestration-cost-layer/. Sources: cursor.com/blog/projects, cursor.com/changelog/projects, cursor.com/docs/cloud-agent (billing: API pricing for the selected model; spend limit required), cursor.com/docs/models-and-pricing — no compute-minute or credits metric exists on the Cursor side; the concurrent-agent count has no published ceiling.

2026-09-14: Rebased the Claude Code Usage Limit Cost Impact Estimator on the permanent 125 index (kanban t_e03f5d47) — the temporary +50% weekly boost ended Sept 13, 2026 at 11:59 PM PT and the permanent level took effect Sept 14, 2026, so every input, default and tooltip that assumed the promotion allowance as the present has been rewritten: new index-basis selector (ccBasis) defaulting to the permanent 125 allowance, with the promotion 150 and pre-promotion 100 bases available for comparison; the usage input now reads % of the selected index basis rather than a percentage of the ended promotion's cap; the seat-price tooltip, the four result cards and both FAQ answers (visible + FAQPage) describe the post-Sept-14 regime. The canonical 100 150 125 index, 125/150 = 83.3% and the 150/125 = 1.2 (+20%) headroom ratio are unchanged and still shown. Promo-current copy removed from the estimator, the FAQ, and the assumptions footer. Verified: node --check, tests/claude-code-limits-verify.mjs, FAQ parity, node harness cases 83.3% / 120% / $600. See the refreshed analysis at Claude Code Usage Limit Cost Impact. 2026-09-12: SEO/AEO pass on the "prices fell, bill rose" pair (kanban t_8988dbaa) — added the exact query as FAQ item 44 on this page (visible item + its FAQPage entry) with the definition (billed tokens × price per token), the arithmetic ($105,600.00 ungoverned against $9,600.00 governed on the same 10,000-task workload, $10.56 against $0.96 per completed task) and the dated anchors ($4.00/$20.00 promotional through November 21, 2026; >272K input tokens repriced at 2x input / 1.5x output for the full request), cited to OpenAI's model and pricing pages and to Bain's September 10, 2026 AI-in-Telecom analysis, and linked to the five governance controls on mybusinessaiaudit.com/ai-agent-risk-checklist#ai-spend-governance. On the why AI bills rise page the section-4 CTA now points at that block instead of the payments-side page (wallet allowances), which never walked the five controls its sentence promised.

2026-09-12: Made the Cache-Aware Agent Runs estimator band-aware (kanban t_fe90b3a6, ruling t_ffdc62d4). The DeepSeek Pricing Basis selector now exposes the published V4 Pro peak band ($0.044 hit / $1.32 miss / $3.96 out per 1M, Mon-Fri 01:00-04:00 and 06:00-10:00 UTC) — off-peak and launch were the only reachable bands while the page’s own copy already quoted a ~6x peak cache-read gap against Claude Fable 5.1’s $0.25/1M (0.25 / 0.044 = 5.68x), so the UI could not produce a figure the prose asserted. The closing sentence of the results note is now band-aware: it cites the rate of the band actually selected instead of printing the off-peak $0.022 on a peak run. The Hybrid option on the Model Strategy dropdown is labelled (neutral baseline, 1.0x) so a preselected default is not read as a recommendation. The site-wide default was left alone by measurement, not preference: flipping it to deepseek-flash would cut the default quote from $2,750/mo to $2,300/mo and raise the default margin from 60% to 68% for custom_agent/small at the markup defaults, which is a pricing recommendation rather than a modelling default (MODEL_COMPUTE_FACTOR.hybrid = 1.0 is the anchor every other factor is defined against). No price, factor or assumption changed and the "Assumptions verified Aug 26, 2026" date is untouched — all three bands were already in CACHE_PRICES. Sources: DeepSeek API pricing (fetched 2026-09-11).

2026-09-12: Stale DeepSeek price sweep (kanban t_3c4b57c8) — audited every DeepSeek price reference on this page against the verified V4.1-Flash fact sheet (t_16f75d78) after the Sept 10 V4.1-Flash launch, and corrected the three that still priced the pre-V4.1-Flash sheet: the "lowest-cost cache-read API" superlative claimed for V4 Pro (the floor is now V4.1-Flash at $0.003/1M off-peak, V4 Pro second at $0.022/1M) in the news entry, the visible FAQ, its FAQPage twin and the model-strategy quote text; and the meta description/keywords block, which carried the Aug 16 V4 era with no V4.1-Flash terms. No retracted claim was introduced: the Sept 14 reroute stays flagged as withdrawn 2026-09-11 and the $0.30/$1.20 pair stays labelled as DeepSeek's peak band. Also linked two live pages from the news list: DeepSeek V4 Pro vs V4.1-Flash pricing: reroute withdrawn, the 100x cached-input spread (previously reachable only from the sitemap) and /orchestration-cost-layer/ (site orphan) — kanban t_9979c663. The two now-false cache-read superlatives inside the dated 2026-08-16 and 2026-08-28 entries below are annotated as superseded rather than rewritten, so the changelog keeps its history without contradicting the current sheet. Sources: DeepSeek API pricing (fetched 2026-09-11) · verified fact sheet t_16f75d78.

2026-09-11: Added the Token Waste & Context Governance factor (#token-governance-estimator — spec kanban t_8813edb2, ship t_1cb93770) and the analysis page Why is my AI bill going up when model prices are falling?. It prices the bill, not the token: billed tokens = useful tokens / (1 − waste), per-call input = context × (calls + 1) / 2 while nothing is pruned, and OpenAI's own rule that prompts above 272K input tokens bill at 2× input / 1.5× output for the full request. Headline, on a held-constant workload of 10,000 tasks a month at 150K context and 8K output per call: $105,600/month ungoverned against $9,600/month governed — 11.0×, or $10.56 against $0.96 per completed task — while the GPT-5.6 Sol rate stays at the promotional $4.00/$20.00 per 1M (a 20% input / 33% output cut, promotional at least through November 21, 2026). On the same window the price effect is −$2,905.26/month and the volume + governance effect is +$96,000.00/month; the long-context cliff alone is $51,200.00 of the ungoverned bill. Model: token-governance-model.js, loaded as its own script, 123 assertions passing; re-read depth and waste rate defaults are illustrative (no vendor publishes an average), every price and the 272K rule are OpenAI's. Umami event token_governance_estimated. Live: publisher check 88 rows compared / 0 mismatches. Sources: OpenAI — GPT-5.6 Sol model page · OpenAI pricing · AWS Bedrock notice (2026-08-21) · Bain (2026-09-10) · IBM (2026-09-10) · Microsoft Azure (2026-09-10) · Unite.AI (2026-09-11) · Consulting.us / McKinsey (2026-09-10).

2026-09-11: Added DeepSeek V4.1-Flash as a selectable model strategy plus a dedicated model-row table and worked example (#deepseek-v41-flash — kanban t_0371431a, from the verified fact sheet t_16f75d78). DeepSeek launched V4.1-Flash on Sept 10, 2026 under the API name deepseek-flash: 552B MoE (8B prefill / 16B decode), native vision, 1M-token context, MIT weights, and off-peak $0.15 input / $0.60 output / $0.003 cached input per 1M with the peak band exactly 2× ($0.30 / $1.20 / $0.006; peak = Mon–Fri 01:00–04:00 and 06:00–10:00 UTC). That is a 100× spread between the peak input rate and the off-peak cached-input rate on the same model — the row exists to show it, so it is rendered from the calculator's own DS41_PRICES data rather than typed into markup. Modeling: MODEL_COMPUTE_FACTOR.ds41flash = 0.84 / MODEL_MARGIN_BONUS.ds41flash = +8 — the cheapest per-token API factor on this page (4.4× below V4 Pro off-peak on cache reads, 5.3× below the previous lowest cache read at $0.016); local keeps the per-workload crown and the #muse-glimmer-local text now states the ranking explicitly. Legacy rows carry visible badges: deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired IDs that are aliased (still accepted, served by V4.1-Flash, billed at the Flash price), and deepseek-v4-pro is NOT retired — the announced 04:00 UTC Sept 14 reroute to V4.1-Flash was WITHDRAWN by DeepSeek on Sept 11, so its row reads "retirement announced 2026-09-10, withdrawn 2026-09-11" and no migration deadline is published. Also updated the two stale "lowest cache-read" claims (the DeepSeek price-increase banner text and the local-vs-API ranking sentence). Sources: DeepSeek API pricing · DeepSeek news — V4.1-Flash launch · HF model card (all fetched 2026-09-11). 2026-09-04: GPT-6 Astra availability refresh — confirmed live in the API and ChatGPT Work/Codex (kanban t_840506ed). OpenAI's developer account confirmed GPT-6 Astra is available in the OpenAI API and in ChatGPT Work and Codex for all Pro, Enterprise, and Business Premium users; ChatGPT shows it as GPT-6 Pro for Pro/Business plans first, Plus chat access to follow. Added Astra Pro tier coverage: Pro/Business/Enterprise plans also get the higher Work/Codex tier, priced separately in Work/Codex credits on top of the included allowance (250 credits per 1M input / 25 cached / 1,250 output; Fast 2.5x in Work/Codex; Codex CLI 0.153.0+); Enterprise admins enable Astra for their workspace — off by default at launch. Rate card unchanged and re-verified Sept 4 (OpenAI model + pricing docs, kingy.ai Sept 4 access matrix, 9to5Mac, HuggingNews). FAQ/schema/helper/copy/meta updated; calculator strategy unchanged (astra 1.32/−8, 272K scenario). Full analysis updated: GPT-6 Astra API Pricing: LIVE Sept 4 — $10/$50 per 1M, 272K Doubling, 1.05M Context. 2026-09-03: Added GPT-6 Astra (OpenAI) as a selectable model strategy — OpenAI's new flagship announced Sept 3, 2026; official API pricing verified the same day: $10 per 1M input / $50 per 1M output / $1 per 1M cached input ($12.50 cache writes). Its 272K rule: prompts with more than 272K input tokens reprice the FULL request at 2x input and cache rates and 1.5x output ($20 input / $75 output / $2 cached / $25 cache writes per 1M). Context window 1,050,000 tokens (922K max input, 128K max output); Batch/Flex 50%, Fast mode 2x. Rollout: Trusted Access / Daybreak orgs Sept 3; Plus/Pro/Business/Enterprise + OpenAI API + AWS in coming days. Modeling: MODEL_COMPUTE_FACTOR.astra = 1.32 / MODEL_MARGIN_BONUS.astra = -8 — above Sol (1.15/-3) and Fable 5.1 (1.25/-6) because Astra's standard tier ($10/$50) is 2.5x Sol's promo input rate and its 272K repricing makes long-context runs the most expensive on the page; OpenAI overhead toggle now also applies to this strategy. Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry, meta description/keywords, Pricing Reference Table footnote, new GPT-6 Astra context scenario (sub-272K vs above-272K interactive cost math), changelog entry. New post: GPT-6 Astra API Pricing: $10/$50 per 1M, 272K Doubling, 1.05M Context. Sources: OpenAI model docs · OpenAI pricing page · OpenAI safety overview (verified Sept 3, 2026). 2026-09-02: Added Gemini 3.8 Flash (Google) as a selectable model strategy and current routing model — released Sept 2, 2026; Google's best reasoning & coding model yet and third Flash-family release in six weeks (3.6 3.7 3.8); intro pricing $0.75/$3.75 per 1M tokens (input/output, output includes thinking tokens) valid through 2026-12-31, then $1.50/$7.50 from Jan 1, 2027 (same structure as 3.7 Flash). 1M-token context, 64K max output; rolling out to Google AI Pro/Ultra (Gemini app, AI Mode, Sheets) and via Gemini API / AI Studio / Antigravity. Google recommends it for software engineering, autonomous agents, and complex multi-step reasoning; notes it can use more tokens at higher effort levels. Cyber variant (3.8 Flash Cyber) is gated via Fairwind — not GA, excluded from client pricing. Modeling kept conservative: MODEL_COMPUTE_FACTOR.gemini38 = 0.93 / MODEL_MARGIN_BONUS.gemini38 = +4 (identical intro to the 3.7 strategy it supersedes — undercuts Qwen 3.8 Max / Grok 4.6 ($2/$6) on both axes during the intro window; rate expires 2026-12-31 and hosted frontier API with no self-host upside). Also added to the Gemini API Cost & Model Routing Savings estimator (model ID gemini-3-8-flash) and the AgentCore estimator model list. Gemini 3.7 Flash rows are preserved and labeled previous version. Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry, meta tags, Pricing Reference Table footnote, changelog entry. New post: Gemini 3.8 Flash API Pricing: Google's Cheapest Frontier Coding Model of 2026. Sources: x.com/Google · Google blog · Model card · Ars Technica · Thurrott · Neowin (verified Sept 2, 2026). 2026-09-01: Updated the Claude Fable 5 model strategy Claude Fable 5.1 (Anthropic GA Sept 1, 2026; API id claude-fable-5-1): same $10/$50 per 1M list but cache reads cut 75% to $0.25/1M (0.025x base input), 1M-token context, 128K max output. Added Claude Mythos 5.1 context: same underlying model with cyber/bio-tuned safeguards, priced identically, trusted access only (CVP/LSVP) — not GA. The 276x cache-gap headline from @JulianGoldieSEO compared Fable 5's $1.00 cache read to DeepSeek's launch $0.003625; with Fable 5.1's $0.25 and DeepSeek's current off-peak $0.022 the gap is ~11x off-peak / ~6x peak — updated in the cache-aware scenario, selector helper, assumption notes, FAQ items + FAQPage schema, meta description/keywords, and changelog. Sources: Anthropic — Claude Fable 5.1 and Mythos 5.1 (Sept 1, 2026) + platform pricing docs (verified Sept 1, 2026). 2026-08-30: Added the Claude Code Usage Limit Cost Impact Estimator — a new section modeling the Sept 14, 2026 weekly limit change: permanent +25% over baseline replaces the temporary +50% boost net −16.7% (≈ −17%) vs today for Pro, Max, Team, and seat-based Enterprise (baseline 100 today 150 Sept 14: 125). Inputs: plan, current weekly usage as % of today's boosted cap (default 100%), seat price ($/mo, default $100), seat count (default 5). Math (research brief t_6a2f4dd0): new cap = today's cap × 125/150 = 0.8333; baseline = today's cap ÷ 1.5; workload at 100% of today's cap = 120% of the new cap (150/125 = 1.2) same workload needs 20% more headroom; cost to keep today's capacity = seat line × 1.2. Outputs: new cap as % of today, usage after Sept 14, capacity cut, dollar cost to hold capacity. Umami event claude_code_limits_estimated fires on submit. Updated the Claude Code weekly-limit FAQ (visible + FAQPage schema) from the Aug 31 promo framing to the Sept 14 schedule and added a new "how much will it cost my agency" FAQ item + schema entry (2830 Q). (Superseded 2026-09-14 — see the 2026-09-14 entry above: the estimator now selects an index basis and defaults to the permanent 125 allowance; the promotion allowance is offered only as an ended comparison.) Meta description/keywords updated. New post: Claude Code Usage Limit Cost Impact. Sources: Anthropic via @ClaudeDevs (Aug 29, 2026) + BleepingComputer + MacObserver + Notebookcheck + usingClaude + AICatchup + Lapaas + Startup Fortune + Help Center promo article (verified via research brief t_6a2f4dd0). 2026-08-29: Added the AgentCore Migration Estimator — a new section modeling the Bedrock Agents Classic AgentCore migration: inputs are agent complexity (simple/managed harness, custom orchestration, multi-agent graph), # agents, sessions/day, active runtime per session (vCPU-hr + GB-hr), tool calls per run ( gateway invocations), memory events per run, long-term memory records, web-search queries/day, policy auth requests, model + tokens per run, and blended agency rate. Outputs: monthly AgentCore platform bill vs Classic-equivalent bill (inference only — Classic is free as a platform [1]), delta/month, one-time migration labor range (hours–1 day / 1–3 weeks / 3–6 weeks [9]), labor cost range, and the "EOL? None — this is a strategy choice" note. Rates are AWS's published AgentCore list prices [3]: runtime $0.0895/vCPU-hr + $0.00945/GB-hr (active only), gateway $0.005/1K, memory $0.25/1K events / $0.75/1K records/mo, web search $7.00/1K, policy $0.000025/request. Umami event agentcore_estimated fires on submit. FAQ item + FAQPage schema entry (2728 Q), meta description/keywords, changelog entry. New post: Bedrock Agents Classic AgentCore Migration Cost. Sources: AWS maintenance-mode doc [1] + AgentCore pricing [3] + DEV.to AWS Builder case study [10] + byteiota labor table [9] (verified via research brief t_bd1e2ee4). 2026-08-29: Added the Wallet allowance-cap breach preset to the Agent Failure & Retry Cost Estimator — models Cloudflare Wallets' manual-override rule: when an agent hits its creator-set spending cap, over-limit requests are blocked at the wallet API layer and routed to a human for a manual override; the approval is billable human time. Math: over-cap requests/mo × (billable minutes per override ÷ 60) × billable rate/hr, added to the retry total with its own "Manual Override Approval" result card. Also labeled the Wallet Fee input in the Agent Wallet & Spend Cap Estimator as ESTIMATE with the flag "Cloudflare has not published wallet fees; this default is an estimate." — no fee schedule exists yet (update when Cloudflare ships fees; tracked in code comment + CHANGELOG). Sources: Cloudflare blog (Aug 4, 2026) + Help Net Security (Aug 5, 2026) — verified via research brief t_7746aecf (Cloudflare agent wallets chain, kanban t_da2044f8). 2026-08-28: Added Qwen3.8-Flash as a selectable model strategy — Qwen's new open-weight multimodal MoE (open-sourced Aug 26, 2026 on Hugging Face + ModelScope; production API on QwenCloud same day): 125B total / ~6B active per token, ~95% sparsity (OrcaRouter phrasing; card says "125B with 6B activated"), 262,144-token native context 1M via YaRN, qwen-community-1.0 license (NOT Apache-2.0). Official pricing: $0.15 per 1M input / $0.47 per 1M output (cache read $0.016); China ¥0.8/¥2.7/¥0.1. OpenRouter live at $0.15/$0.47; Artificial Analysis medians $0.30/$1.20; OrcaRouter pass-through. Modeling: MODEL_COMPUTE_FACTOR.qwen38flash = 0.86 / MODEL_MARGIN_BONUS.qwen38flash = +7 — the cheapest open-weight API at list (below GLM-5.3-Flash on output and cache, below DeepSeek V4 Pro off-peak on both axes, lowest cache-read on the page at $0.016 at the time — superseded 2026-09-10 by DeepSeek V4.1-Flash at $0.003/1M off-peak), tied with dsv4pro's 0.86 rather than below because benchmarks are vendor-reported with no independent replication, the model is two days old, GLM-5.3-Flash's 50% promo still undercuts on raw price through Sep 9, and the qwen-community-1.0 license is a new licensing layer. Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry (2425 Q), meta description/keywords, Pricing Reference Table footnote, open-weight section narrative + sources, changelog entry. New post: Qwen3.8-Flash-Next Cost: Open-Weight MoE at $0.15/M. Sources: QwenCloud qwen3.8-flash + HF Qwen/Qwen3.8-Flash-Next + OpenRouter qwen/qwen3.8-flash + OrcaRouter release blog + Artificial Analysis + Unite.AI + Intelligent Living + SGLang docs + HN thread (verified via kanban t_ca417b6f / t_d363cc52). 2026-08-28: Added OpenAI's GPT-5.6 discount-experiment data to the model-cost comparison pages: during Jul 27 – Aug 14, 2026, Terra and Luna were discounted 50% via OpenRouter — daily Terra token usage rose 5.6x and Luna 13.8x, while Sol at list price rose only 1.1x (control); ~1/3 of users who tried a discounted model kept using it after expiry, and Terra/Luna went from 0.7% to 7.8% of OpenRouter tokens (Jevons-paradox elasticity, per OpenRouter's Aug 25 analysis + TLDR Aug 28). List prices updated: GPT-5.6 Terra $2/$12 (−20% Jul 30, from $2.50/$15), GPT-5.6 Luna $0.20/$1.20 (−80% Jul 30, from $1/$6) per 1M tokens — added to the per-1M rate tables on AI Model Cost per Task 2026 and AI Agent API Cost Calculator, plus the Sol pricing page. Sol's $4/$20 promo (Aug 21 – Nov 21, 2026) unchanged. New post: OpenAI's Discount Experiment: What 13.8x Usage Growth Means for Agency Pricing. Sources: OpenRouter Blog (Aug 25, 2026) · TLDR AI (Aug 28, 2026) · OpenAI API pricing (platform docs, verified Aug 28, 2026) — research brief t_5c8841ca (7 sources, 18 evidence quotes). No calculator-logic change. 2026-08-28: Added a PRICING-PATH ALERT + FAQ entry: OpenRouter removed all 35 openai/*:batch model slugs on Aug 26, 2026 (plus z-ai/glm-5.2:batch and moonshotai/kimi-k2.7-code:batch) — catalogue went 422380 models, 37 of 42 departures carried the :batch suffix. The slugs still answer metadata requests, so nothing looks broken until a completion returns 404 "No endpoints found". Base models untouched (openai/gpt-5 still routes at $1.25/$10.00 per 1M; OpenAI vendor page: 96 models, ~2.8T tokens/wk); 24 batch variants from Anthropic (11), Google (10), Thinking Machines/NVIDIA/MiniMax survive — anthropic/claude-opus-5:batch still bills at the clean 50%. No calculator-logic change (batch is a routing tier, not a model price input), but quoting guidance updated: (1) OpenRouter beta Batch API — POST /api/beta/batches, plain slug, 50% of standard per-token pricing, 24h window, text-only (rejects image/audio/file parts), returns 202 + status "validating" (do not treat 202 as completion); (2) openai/flex — provider: {"only": ["openai/flex"]}, synchronous 50% (gpt-5.6-sol $1.00/$5.00 vs $2.00/$10.00), explicit opt-in required, GPT-5.6 family only (absent on gpt-5, gpt-4o, o3, gpt-4.1); (3) OpenAI direct Batch API — still 50% vs synchronous, separate rate-limit pool, 50K requests/batch, 200MB input, 2,000 batches/hr, completes within 24h. Check any slug: curl .../api/v1/models/<model>/endpoints — empty array = dead; run it against every model string in config (42 entries vanished). FAQ item + FAQPage schema entry (2324 Q), visible FAQ item, model-pricing NEW banner, meta description/keywords, changelog entry. New post: OpenRouter OpenAI Batch Endpoints Down: 3 Ways to Keep the 50% Discount. Sources: MoClaw investigation (Aug 28, 2026) + OpenRouter /api/v1/models live read + OpenRouter Batch API Quickstart + provider routing docs + OpenAI Batch API docs (verified via kanban t_40c4dca7). 2026-08-28: Added Tencent Hy4 preview as a selectable model strategy — Tencent's largest open-weight release yet (open-sourced Aug 28, 2026): 770B total / 49B active MoE, native 1M-token context, 64K max output, Apache 2.0 weights on Hugging Face (tencent/Hy4-preview + FP8). OpenRouter lists $0.834 per 1M input / $2.501 per 1M output (cached input $0.042), Tencent Cloud sole provider — ~40% below GLM-5.3 on input, ~43% below on output, ~6x below Kimi K3 on output, still above DeepSeek V4 Pro off-peak ($0.66/$1.98). Tencent's internal blind eval (163 experts, 203 tasks): 2.99/4 vs GLM-5.3 2.92 / Kimi K3 2.94 — tier parity, not decisive. Modeling: MODEL_COMPUTE_FACTOR.hy4preview = 0.90 / MODEL_MARGIN_BONUS.hy4preview = +5 — cheaper per-token than the plain open stack at list, kept above glm53flash's 0.88 because GLM-5.3-Flash's promo/rate still wins on raw price and DeepSeek keeps the cache floor; flagged in output. Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry (2223 Q), meta description/keywords, Pricing Reference Table footnote, open-weight section narrative + sources, changelog entry. New post: Tencent Hy4 Preview Pricing: 770B Open Weights at $0.834/M. Sources: Tencent release + hy.tencent.ai + GitHub Tencent-Hunyuan/Hy4-preview + Hugging Face tencent/Hy4-preview + OpenRouter tencent/hy4-preview + Bloomberg (Aug 28, 2026) + KuCoin flash + explainx.ai (verified via kanban t_6e671c22). 2026-08-26: Updated the ChatGPT Ads (EU) pricing preset with OpenAI's enterprise-first trajectory (Digiday, Aug 24, 2026): OpenAI is hiring a head of ads enterprise marketing ($374K–$415K + equity) to position ChatGPT Ads for enterprise advertisers and agencies; Ads Manager tooling now targets media teams (custom audiences need a 25K minimum, account-health indicators, budget alerts, CMO-ready reporting charts, automated headline/copy suggestions) and an SMB team is in build-out — enterprise/agency-led budgets scale first, SMB self-serve later. Pricing implication: enterprise engagements can command the premium end of the setup/management ranges now; keep SMB quotes as instrumented pilots. Updated: preset assumptions box, FAQ item + FAQPage schema entry (2122 Q), meta description, changelog entry. Source: Digiday — OpenAI is sharpening its focus on enterprise advertisers (verified via kanban t_4bc87ea6). 2026-08-26: Added GLM-5.3-Flash (Z.ai) as a selectable model strategy — the anonymous Ox Alpha stealth model (free on OpenRouter/OpenCode from Aug 20) was confirmed by Z.ai on Aug 26, 2026 as GLM-5.3-Flash: 320B/18B MoE, natively multimodal, 1M-token context, MIT weights live on Hugging Face, served entirely on Chinese AI chips during the preview. Official API pricing: $0.15 per 1M input / $0.50 per 1M output (cached input $0.03), with a 50% launch promo to $0.075/$0.25 through Sep 9, 2026 — ~1/10th of the flagship GLM-5.3 rate ($1.40/$4.40) and below DeepSeek V4 Pro off-peak ($0.66/$1.98) on both axes. Modeling: MODEL_COMPUTE_FACTOR.glm53flash = 0.88 / MODEL_MARGIN_BONUS.glm53flash = +7 — cheaper per-token than the plain deepseek stack at list, kept above dsv4pro's 0.86 because the cache-aware profile still wins long agent runs on cache reads, benchmarks are vendor-reported (DeepSWE v1.1 63.4, TB2.1 84.3, AutomationBench 48.8, Code Bench 29.0 vs Opus 4.8 29.5), and the promo rate doubles after Sep 9. Also replaced the 2 stale "GLM 5.3 unconfirmed" passages (open-weight section + FAQ) with the confirmed reveal. Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry (2021 Q), meta description/keywords, changelog entry. New post: GLM-5.3-Flash Pricing: Ox Alpha Revealed at $0.15/M. Sources: Z.ai release + Bloomberg/TechCrunch (Aug 26, 2026) · Kingy price check (Aug 26) · Wccftech · Hugging Face zai-org/GLM-5.3-Flash (verified via kanban t_f9df0f69). 2026-08-26: Added Claude Opus 5 (Anthropic) as a selectable model strategy and added Ramp AI Index adoption/enterprise-share stats to the Fable 5, Opus 5, and GPT-5.6 Sol entries: Opus 5 — $5 per 1M input / $25 per 1M output (exactly half of Fable 5; Anthropic official, launched late July 2026), modeled MODEL_COMPUTE_FACTOR.opus5 = 1.18 / MODEL_MARGIN_BONUS.opus5 = -5 (between Sol 1.15/-3 and Fable 5 1.25/-6); Fable 5 — ~6% of Anthropic business tokens / 11.4% of Anthropic spend (July 2026); GPT-5.6 Sol — ~25% of OpenAI tokens / 23% of OpenAI spend; FT-cited Ramp data reports Opus 5 overtook Fable 5 in enterprise spend after its late-July debut (attributed to FT-relayed Ramp data via Superpower Daily/CIO — WinBuzzer/BreezyScroll do not confirm the overtake). Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry, meta description/keywords, Pricing Reference Table footnote, assumptions footer (date Aug 26), changelog entry. Sources: Ramp AI Index Aug 2026 · WinBuzzer · BreezyScroll · Superpower Daily (FT) · Anthropic official (verified via parent research brief t_06f1cfb2). 2026-08-25: Added a Meta Hatch consumer AI agent comparison row — new Consumer AI Agent Pricing section with an 8-field comparison table (provider, product, price, billing, availability, model, category, source). Meta Hatch: $199.99/mo premium tier — REPORTED, pending confirmation (The Information Jun 4 + Aug 24, 2026, via PYMNTS/RuntimeWire; final pricing NOT set; not shipped as of Aug 25, 2026), monthly billing, launching in coming weeks (late Aug–early Sep 2026 target), Watermelon model targeted for October 2026, consumer AI agent category — compared against ChatGPT Plus ($20/mo), ChatGPT Pro ($100–$200/mo) and Claude Max (up to $200/mo). Price-update mechanism documented in an HTML comment + site CHANGELOG; analytics task t_ecfe7f64 monitors sources for official pricing. FAQ item + FAQPage schema entry, meta description/keywords, assumptions footer note, changelog entry. Full explainer: Meta Hatch AI Agent Price: What Consumer AI Agents Cost (verified via parent research t_b6570320). 2026-08-25: Added the ChatGPT Business Seat Cost Estimator — a new section modeling Standard vs Premium seats with OpenAI's verified list prices (Premium $125/user/mo monthly / $100 annual (20% discount); Standard $25/$20), billing cadence (monthly vs annual), seat count (2-seat min / 200-seat cap), and the usage-multiplier assumptions (Premium = 5x more usage than Standard, no five-hour limit, weekly resets). Example: 10 Premium seats billed annually = $1,000/month ($12,000/yr); billed monthly = $1,250/month. Premium seats are now a calculator input — the Aug 10/12 "budget separately" note is superseded. Verified via parent research brief t_642ada39 (7 sources, 34 evidence quotes; announced August 2026, official launch Aug 25, 2026). FAQ item + FAQPage schema entry, meta tags, assumptions footer (date Aug 25), changelog entry. New explainer: ChatGPT Business Premium Seats Pricing (2026). 2026-08-23: Added a Codex 20M users / banked reset note + FAQ entry: Codex and ChatGPT Work crossed 20 million active users the week of Aug 21, 2026 (up from 15M a week earlier), and OpenAI credited every paid user a banked reset (saved usage-limit credit that resets both 5-hour and weekly windows when redeemed). The reset rollout missed its Aug 21 deadline; OpenAI set a firmer 8pm PST same-evening deadline that also passed for many users, then announced a new reset for 2026-08-24T21:00:00Z (2pm PST) after finding usage inefficiencies. No calculator-logic change and no price change — plan list prices stand (Plus $20, Pro 5x $100, Pro 20x $200, Business per seat, API usage-based); the milestone is a capacity/billing-clarity story, not a price input. FAQ item + FAQPage schema entry, meta description/keywords, changelog entry. New explainer: Codex Pricing 2026: Plans, Limits & Banked Reset for Agencies (companion: Codex Just Hit 20 Million Users — Find AI Agency; verified via research brief t_72fe5cf1). 2026-08-22: New post GPT-5.6 Sol API Pricing: OpenAI Cuts Rates to $4/$20 (Aug 21 – Nov 21, 2026) — first Sol price cut since launch, effective Aug 21 through at least Nov 21, 2026. 2026-08-21: New explainer AI Client Communication Workflow: What It Costs to Automate Client Texts — OpenAI's Apple Messages plugin (Aug 20, 2026) lets ChatGPT read, search, summarize, draft, and send iMessage/SMS/RCS on Apple silicon Macs in ChatGPT Work/Codex. The plugin itself is plan-inclusive; the real cost line items are hardware (Apple silicon per seat), Work/Codex seats priced into retainer math, and compliance review (OpenAI has not documented exactly which message content leaves the machine or published managed-Mac admin guidance). No calculator-logic change — this is a client-communication workflow cost lens, not a token-price input. Sources: OpenAI release notes (Aug 20, 2026) + OpenAI Codex plugin docs + 9to5Mac/Engadget/MacRumors/TechCrunch/TNW/Unite.AI/Yahoo (verified via research brief t_c7d9dc7e). 2026-08-22: GPT-5.6 Sol API pricing updated to the verified promo rate — OpenAI published official per-token pricing for GPT-5.6 Sol on Aug 21, 2026: $4 per 1M input / $20 per 1M output / $0.40 per 1M cached input, a promotional rate valid Aug 21 – Nov 21, 2026, down from the $5/$30 estimate the calculator previously used (API + eligible ChatGPT Work/Codex credits; ChatGPT Pro/Plus/Business subscription pricing unchanged). All Sol pricing references updated to $4/$20 with the promo window: selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entries, meta description/keywords, Pricing Reference Table footnote, assumptions footer (added Sol promo note; Grok 4.6 comparison updated from 60%/80% cheaper to 50%/70% cheaper vs $4/$20), changelog entry. No other model prices changed. Calculator modeling unchanged (MODEL_COMPUTE_FACTOR.sol = 1.15 / MODEL_MARGIN_BONUS.sol = -3 — still premium tier). Cost-per-task verification: 10K-in/2K-out task drops from ~$0.110 to ~$0.080 (-27%); 100K-in/5K-out drops ~$0.650 to ~$0.500 (-23%); 20K-in/4K-out drops ~$0.220 to ~$0.160 (-27%) — input-heavy agent workloads drop ~20%+, consistent with the 20% input / 33% output cuts. Sources: OpenAI API pricing (published Aug 21, 2026), verified via research brief t_1d150dd3 (7 sources, 16 evidence quotes). 2026-08-22: Added a ChatGPT Ads (EU) pricing preset — new Service Type option " ChatGPT Ads Management (EU)" with labeled inputs (monthly ad spend, management fee %) shown when selected, plus a sourced assumptions box. OpenAI expands ChatGPT Ads to 31 European markets Aug 24, 2026 (40 total); ads run on Free + Go tiers only (Go ≈ €8/mo); buying is agency-led first (OpenAI Ads Solutions + agency/tech partners; self-serve Ads Manager later this summer). Pricing model: setup $1,500–$5,000 + management fee 15–25% of ad spend, floored at $2,000/mo and capped at $7,500/mo (ad spend is pass-through — client pays OpenAI; the fee is the agency's revenue). Margin 62% base; portability/plugin factors intentionally skipped for this managed-service line. FAQ item + FAQPage schema entry, meta description/keywords, Pricing Reference Table row, assumptions footer (date Aug 22), changelog entry. Sources: OpenAI (Aug 18, 2026) · Search Engine Land · Dataconomy · Euronews (verified via research brief t_ff98639f). 2026-08-20: Added a Claude Code 50% weekly usage-limit boost note + FAQ entry: Anthropic extended the promo through August 31, 2026 (11:59 PM PT) — the third extension since the May 13, 2026 launch — and for the first time said it hopes to make the boost permanent while warning capacity may be tight (announced Aug 18, 2026 via @ClaudeDevs; widely reported Aug 19). Boost applies automatically to Pro, Max (5x/20x), Team, and legacy seat-based Enterprise; excludes Free and consumption-based Enterprise seats; Claude Code only (CLI, IDE, desktop, web) — 5-hour session limits, Claude chat, Claude Cowork unchanged; weekly limits "return to their standard levels" after Aug 31. No calculator-logic change — the boost is a seat-plan weekly-quota layer, and this calculator models API/token and project pricing, not seat-based subscriptions. FAQ item + FAQPage schema entry, changelog entry. Example math: Max 5x at $100/mo delivers 1.5× weekly quota through Aug 31 ≈ $150/mo equivalent at standard limits. (Superseded 2026-09-14 — see the 2026-09-14 entry above: the promotion ran May 13 through Sept 13, 2026 and the permanent +25%-over-baseline level took effect Sept 14, 2026.) Sources: Anthropic Help Center — Claude Code May–August 2026 weekly limits promotion · @ClaudeDevs (Aug 18, 2026) (verified via research brief t_4e969eaa). 2026-08-20: Added the OpenAI safety-monitoring overhead stress-test toggle — a visible checkbox in the Model Strategy section (OFF by default) that applies a +20% multiplier to OpenAI-based strategies (GPT-5.6 Sol, Paid frontier). Basis (verified via research brief t_b34197c5): OpenAI's official post (Aug 18, 2026) estimates its expanded chain-of-thought safety monitoring adds roughly 20% overhead to the inference compute it monitors — required for all RL training/eval involving tools for models of Sol capability or higher, plus all Astra inference with tools after the Aug 7 Critical-cyber determination; OpenAI is NOT currently billing customers for this overhead (spokesperson via The Register, Aug 19, 2026; corroborated TNW — not stated in the official post), and Anthropic says its safeguards make a similar slowdown unnecessary (Axios, Aug 19, 2026). The toggle is a what-if cost stress, not a price change: no default outputs change (OFF by default); the output note explains when it applies vs. not; FAQ item + FAQPage schema entry, meta description/keywords, assumptions footer (date Aug 20), changelog entry. Sources: OpenAI (Aug 18, 2026) · The Register (Aug 19, 2026) · TNW (Aug 19, 2026) · Axios (Aug 19, 2026). Updated Aug 20 (publish): new explainer OpenAI Inference Overhead: The 20% That Matters for AI Agency Pricing — the agency-facing read on why this is a cost-side signal (not a price input), what's in scope (Sol-class tool-enabled RL training/eval; Astra inference with tools), and how to stress-test frontier assumptions with the toggle. 2026-08-19: Added an AI supplier risk note + FAQ entry: on Aug 18, 2026 OpenAI announced a two-week pause in RL training on its latest deployment-bound models and held its largest planned frontier RL run on hold after internal evaluations flagged its upcoming Astra model at the "Critical" cyber-capability threshold under its Preparedness Framework (core training continued; no model cancelled; no API price changes — calculator math unchanged). Guidance: include fallback AI models and review provider tooling roadmaps when evaluating AI costs/risks. FAQ item + FAQPage schema entry, assumptions footer, changelog entry. Source: OpenAI (Aug 18, 2026) (verified via research brief t_bb45d771). Updated Aug 19 (eve): supplier-risk FAQ now also notes OpenAI's committed 2027 public listing (CFO Sarah Friar, CNBC Aug 19) de-risks vendor longevity on a known timeline; new explainer OpenAI's 2027 IPO Window: Re-Baseline Your Cost Assumptions. Updated Aug 20: Anthropic now expects to match or top SpaceX's record IPO (~$75B outset / $86.2B w/ overallotment) and could file by end of August at a valuation just under $1T (Bloomberg Aug 20) — the pre-IPO window is the Claude pricing-risk zone; new explainer Anthropic's Record IPO: Stress-Test Your Claude Cost Assumptions. 2026-08-16: Added DeepSeek V4 Pro (cache-aware) and Claude Fable 5 as selectable model strategies, plus a dedicated Cache-Aware Agent Runs cost scenario (#deepseek-cache-scenario) — verified via parent research brief t_40deadfe. DeepSeek's official peak/off-peak pricing LANDED 2026-08-16 16:00 UTC (off-peak V4 Pro: cache hit $0.022/1M, miss $0.66/1M, output $1.98/1M; peak: $0.044/$1.32/$3.96 — DeepSeek pricing page + news260813), superseding the PROVISIONAL pre-hike schedule. Claude Fable 5 (Anthropic, claude-fable-5): $10/$50 per 1M with $1.00/1M cache reads — the premium comparison. Modeling: MODEL_COMPUTE_FACTOR.dsv4pro = 0.86 / MODEL_MARGIN_BONUS.dsv4pro = +7 (cache-aware V4 Pro was the cheapest cache-read stack on this page for agent runs — 92% of input bills at $0.022/1M — but kept below the plain deepseek factor because the 92% hit rate is an assumption, flagged in output; superseded 2026-09-10 by DeepSeek V4.1-Flash at $0.003/1M off-peak — see the 2026-09-12 entry above); MODEL_COMPUTE_FACTOR.fable5 = 1.25 / MODEL_MARGIN_BONUS.fable5 = -6 (most expensive per-token stack on the page). The ~276x cache-read gap from @JulianGoldieSEO's Aug 16 post is arithmetically correct on DeepSeek's LAUNCH pricing ($1.00/$0.003625 = 275.9x) but STALE after the increase — current gap ~45x off-peak / ~23x peak, and the 92% hit rate is an author assumption, not a published OpenRouter statistic (both caveats shown prominently in the scenario, FAQ, and assumptions). Selector options + helpers, model labels + assumption notes + ROI text, FAQ item + FAQPage schema entry (new cache-gap Q&A, DeepSeek Q&A rewritten from 'provisional' to 'landed'), meta description/keywords, Pricing Reference Table footnote, open-weight section banner, assumptions footer (date Aug 16, 2026), changelog entry. Sources: DeepSeek pricing · DeepSeek release note · Anthropic pricing · Julian Goldie SEO full post (verified via research brief t_40deadfe). 2026-08-13: Added Claude Sonnet 5 (Anthropic) as a selectable model strategy — $2 per 1M input tokens / $10 per 1M output tokens, PERMANENT as of Aug 10, 2026. Anthropic made the launch pricing permanent and cancelled the previously scheduled Sept 1, 2026 increase to $3/$15 — no scheduled price change is pending (verified Aug 10, 2026 announcement + official Claude pricing docs; research brief t_735f2ef6). Cache: $0.20 per 1M cached input; cache write $2.50 (5m) / $4 (1h) per 1M. Modeling kept conservative: MODEL_COMPUTE_FACTOR.sonnet5 = 0.97 and MODEL_MARGIN_BONUS.sonnet5 = +3 — input at open-weight parity (Qwen 3.8 Max $2/$6), output above open-weight ($10 vs $6) but well below premium frontier (GPT-5.6 Sol $5/$30 estimate), hosted frontier API with no open-weight/self-host upside. Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry, meta tags, Pricing Reference Table footnote, assumptions footer, changelog entry. Sources: Anthropic announcement · Claude pricing docs (verified via research brief t_735f2ef6). 2026-08-13: Added Gemini 3.7 Flash (Google) as a selectable model strategy — released Aug 13, 2026; intro pricing $0.75/$3.75 per 1M tokens (input/output, output includes thinking tokens) valid through 2026-12-31, then $1.50/$7.50 from Jan 1, 2027 (post-intro equals 3.6 Flash's launch price; intro is exactly half). 1M-token context, 64K max output, free tier, 5,000 free search requests/mo shared across Gemini 3.x, 50% batch discount. vs Gemini 3.6 Flash: FrontierCode 1.1 Main 43.6% (vs 34.4%), DeepSWE v1.1 65.3% (vs 49.0%), WebDev Arena Elo 1588 (vs 1538). Modeling kept conservative: MODEL_COMPUTE_FACTOR.gemini37 = 0.93 (slightly below open-weight 0.95) and MODEL_MARGIN_BONUS.gemini37 = +4 — intro pricing undercuts Qwen 3.8 Max / Grok 4.6 ($2/$6) on both axes, but the rate expires 2026-12-31 and post-intro output ($7.50) is above open-weight ($6), and it is a hosted frontier API with no open-weight/self-host upside. Also added Gemini 3.7 Flash to the Gemini API Cost & Model Routing Savings estimator as a routable current model (model ID gemini-3-7-flash). Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry, meta tags, Pricing Reference Table footnote, assumptions footer (date Aug 13, 2026), changelog entry. Sources: Google blog · Gemini API pricing · Model card (verified via research brief t_7ea3f430). 2026-08-12: Qwen 3.8 Max open weights landed on Hugging Face (Qwen/Qwen3.8-2.4T-A95B — 2.4T params / 95B active — plus Qwen3.8-2.4T-A95B-FP8, Aug 12 ~10:24 UTC) with the official Qwen3.8-Max License. Replaced the monitor pending final terms alert with the finalized terms: separate Qwen license required before commercial use ONLY if you/affiliates run a Model-as-a-Service or AI Work Assistant business AND aggregate revenue exceeds US$50M over any consecutive 12 months (2.5x higher than Kimi K3's $20M/12mo); the NEW AI Work Assistant trigger (independent AI product primarily for AI-assisted coding or office productivity — e.g. Qoder, QwenWork; excludes single-purpose tools, non-coding/office assistants, and assistants that are a feature of another product) is absent from Kimi K3; MaaS = third-party inference/fine-tuning access with meaningful control over inputs/params/training data (mere relaying to third-party-hosted models excluded); internal use not exposed to third parties is exempt; no numeric % rate in the license — the separate license is individually negotiated (Reuters "up to 30%" is reporting, not license text); display clause identical to Kimi K3 (100M MAU or $20M/month revenue prominently display the model name). Updated: alert box, selector helper, Pricing Reference Table footnote, FAQ item + FAQPage schema entry, meta tags, open-weight section narrative (weights now downloadable), assumptions footer, changelog. No calculator-logic change — the $2/$6 per 1M list price stands; the license is a threshold-based commercial-licensing layer. Sources: HF model card · official LICENSE · FP8 weights (verified live 2026-08-12). 2026-08-12: Added Grok 4.6 (SpaceXAI) as a selectable model strategy — frontier API at open-weight prices: released Aug 12, 2026; $2/$6 per 1M tokens (input/output), fast variant 2x, cache hit $0.50 per 1M input (-75%, AA); 500k context and Intelligence Index 61 (= GPT-5.6 Sol max) per Artificial Analysis (x.ai announcement silent on context); vs GPT-5.6 Sol $5/$30 (AA), 60%+ cheaper on input, 80% cheaper on output. Modeling kept conservative: MODEL_COMPUTE_FACTOR.grok46 = 1.0 (neutral vs hybrid baseline) and MODEL_MARGIN_BONUS.grok46 = +2 — the verified list price matches the open-weight tier (Qwen 3.8 Max $2/$6) but it is a hosted frontier API with no open-weight licensing/self-host upside. Selector option + helper, model label + assumption note + ROI text, FAQ item + FAQPage schema entry, meta tags, Pricing Reference Table footnote, assumptions note, changelog entry. Availability: Cursor, Grok Build, SpaceXAI API, OpenRouter, Vercel, Cloudflare (sources: x.ai/news/grok-4-6, 9to5Mac, Artificial Analysis; verified via research brief t_a87cfcc9). 2026-08-12: Added Grok Bot (SpaceXAI) as the fourth AI coding agent billing model — subscription-bundled agent access — in the cost-transparency comparison: no standalone price announced (unknown / contact sales placeholder), bundled with SuperGrok Heavy, Cursor Ultra, Cursor Teams Premium; always-on agents run 24/7 on their own cloud computer; desktop (macOS) + iOS; enterprise waitlist; no usage caps disclosed. Agent-option metadata block, FAQ item + FAQPage schema entry, meta tags, and assumptions note updated. No calculator logic change — Grok Bot is not usage-metered, so it is not a calculator input (sources: x.ai/news/introducing-grok-bot, MacRumors, Oflight; verified via research brief t_e203f6d1). 2026-08-12: Fleshed out the Local / self-hosted (Meta Muse Glimmer) option with the official Ollama runtime and a same-model hosted benchmark: ollama run muse-glimmer (official 18 GB build / 128K context / text+image; Apple Silicon MLX build muse-glimmer:30b-mlx, 21 GB), ollama launch integrations (Claude Code, Codex, Pi, OpenCode, GitHub Copilot, OpenClaw, Hermes Agent). Meta has published no official hosted API price, so the scenario now benchmarks against Together AI hosting the same 30B at $0.35 / $1.50 per 1M tokens ($0.04 cached input) — verified on Together's pricing and model pages (2026-08-12) — making the local-vs-API break-even concrete for the same model, not just vs other vendors. Selector label, helper text, verified-specs table (new "Local runtime (Ollama)" row), cost-scenario narrative, FAQ answer, source list (+2), and assumptions date (Aug 12) all updated. 2026-08-11: Added Local / self-hosted (Meta Muse Glimmer) as a selectable model strategy — the cheapest compute factor on this page (0.85× vs open 0.95×, margin +8) — reflecting that the cheapest-model answer is no longer just "which API": Meta's Muse Glimmer (released Aug 10, 2026; 30B dense, Apache 2.0, 131,072+ context, official 4-bit K-Quant-17GB GGUF = 16,756,681,056 bytes / ~15.6 GiB / under 20 GB, single consumer GPU at >200 tok/s on RTX 5090 with DFlash) runs at electricity-cost marginal inference once a 24 GB-class card is amortized. New dedicated Local vs API cost scenario section (#muse-glimmer-local) with verified specs table, break-even framing vs hosted APIs (DeepSeek V4 PROVISIONAL $0.14/$0.28, Qwen 3.8 Max $2/$6, Kimi K3 $3/$15), honest caveats (1.0% quant degradation, 24–32 GB envelope for full context, not frontier on HLE/GPQA, no audio, you own ops/security), and source links (Meta model card, GGUF repo, HF blog, NVIDIA blog, TechStartups, AMD, explainx). Model Strategy selector helper, open-weight section narrative, Pricing Reference Table footnote, FAQ item + FAQPage schema entry, meta tags, and assumptions date (Aug 11) all updated. Per-workload framing — "cheapest per workload," not blanket "cheapest model."

2026-08-08: Added DeepSeek V4 (V4-Flash / V4-Pro) as a selectable model strategy with a PROVISIONAL pricing flag — DeepSeek announced (Aug 6, 2026, footnote 2 of its official pricing page) that it plans to raise overall API pricing "in the near future, with a significant increase expected," and has NOT disclosed new rates, a percentage, or an effective date as of 2026-08-08. Current list prices (V4-Flash $0.14/$0.28, V4-Pro $0.435/$0.87 per 1M tokens) are treated as pre-hike/provisional: the strategy is deliberately given a smaller discount than the pure open-weight stack so the calculator does not over-state "DeepSeek is cheapest," and the output carries an explicit verify-before-quoting note. New prominent red warning banner in the Open-Weight section, FAQ item + FAQPage schema entry, meta tags extended, sources linked (DeepSeek pricing page live 2026-08-08, TNW, SCMP, Dataconomy/Bloomberg, TechNode). Open-weight self-host escape hatch noted (MIT weights).

2026-08-08: Added a monitor pending final terms alert for Qwen 3.8 Max: Alibaba (Reuters, Aug 7) plans a revenue-share requirement for qualifying model-as-a-service use (rates TBD, effective date TBD) — the cost model no longer presents Qwen 3.8 Max as free/unrestricted at scale. Kimi K3 precedent verified at >$20M annual sales for a commercial agreement (up to 30% revenue share). New FAQ item + FAQPage schema entry; meta tags extended. Source: Reuters Aug 7, 2026.

2026-08-06: Added Qwen 3.8 Max to the open-weight model strategy (GA Aug 2–3, 2026; 2.4T-param MoE, ~95B active; $2/$6 per 1M tokens; 1M context; open weights promised ~Aug 10). Real-world usage section added with the Aug 6, 2026 45-project field report — self-reported, no artifacts; the "destroyed Fable 5" claim was walked back by the tester (Qwen strong on fast/multimodal builds; Fable 5 on huge long-running projects). Transparent cost-variability note added: API list prices are public, but self-hosted cost is hardware/quantization-dependent and not yet knowable until weights drop.

2026-08-06: Added GPT-5.6 Sol as a selectable model strategy (frontier, Instant + deep reasoning) with a clearly marked estimate and link to OpenAI's pricing page — OpenAI announced Sol now powers both Instant and deep reasoning for Plus/Pro, and GPT-5.6 Luna becomes the default for Free/Go users (unlimited text chats rolling out this week/next week). No official per-token API pricing published for either model; assumptions date/source now included in calculator output.

2026-08-05: Added Gemini API cost & model routing savings estimator and explainer (Google Cloud managed model routing, Public Preview Aug 3, 2026). Pricing sourced from Google's published Vertex AI / API Gateway list prices; scenario figures are illustrative (directional).

Open-Weight Models: The New Cost Lever for Agencies

✓ Qwen 3.8 Max license terms FINALIZED — open weights live on Hugging Face (Aug 12, 2026)

Qwen 3.8 Max's official license requires a separate Qwen license before commercial use ONLY if you (or an affiliate) run a "Model as a Service" or "AI Work Assistant" business AND aggregate revenue exceeds US$50,000,000 over any consecutive 12 months. There is no numeric % rate in the license — the separate license is individually negotiated (Reuters' "up to 30%" is reporting, not license text). "AI Work Assistant" = an independent AI product primarily for AI-assisted coding or office productivity (e.g. Qoder, QwenWork); it excludes single-purpose tools, non-coding/office assistants, and assistants that are merely a feature of another product. MaaS = giving third parties access to inference/fine-tuning (API or hosted endpoint) with meaningful control over inputs, parameters, or training data; merely relaying to third-party-hosted models is excluded. Internal use not exposed to third parties is exempt. A separate display clause (identical to Kimi K3) requires prominently showing the model name on any commercial product/service UI above 100M monthly active users or $20M/month revenue. For most agencies — below the $50M/12mo bar and not running a public MaaS or AI Work Assistant product — the $2/$6 per 1M token list price and self-hosted math stand unchanged.

Sources: HF model card — Qwen3.8-2.4T-A95B · official Qwen3.8-Max License (verbatim) · FP8 weights · Reuters — Alibaba revenue-share plan (Aug 7, 2026, background)

DeepSeek price increase LANDED — official peak/off-peak rates effective 2026-08-16 16:00 UTC

DeepSeek's announced price increase is now in effect. Official rates (DeepSeek pricing page, live 2026-08-16): off-peak V4 Pro cache hit $0.022 / cache miss $0.66 / output $1.98 per 1M tokens; peak $0.044 / $1.32 / $3.96. The pre-hike schedule (V4-Flash $0.14/$0.28, V4-Pro $0.435/$0.87 per 1M tokens) is superseded. Even at the new rates, DeepSeek V4 Pro keeps a bottom-tier cache-read position for long agent runs — cache reads at $0.022/1M off-peak, now second to DeepSeek V4.1-Flash at $0.003/1M off-peak (deepseek-flash, launched Sept 10, 2026 — see the V4.1-Flash row, 100× spread and worked example ; V4 Pro itself is NOT retired — the Sept 14 reroute was withdrawn Sept 11) vs Claude Fable 5.1's $0.25/1M (~11x off-peak gap; the widely-shared "276x" figure compared Fable 5's $1.00 to DeepSeek's launch pricing, $1.00/$0.003625, and is stale on both sides after the increase and Fable 5.1's Sept 1 cache-read cut). The calculator's DeepSeek V4 strategy now reflects the landed rates and a new DeepSeek V4 Pro (cache-aware) strategy models agent-run economics at a 92% hit-rate assumption (see the cache-aware scenario ). Re-verify api-docs.deepseek.com/quick_start/pricing before quoting. Self-hosting remains an escape hatch: V4 weights are MIT-licensed open weights (1.6T Pro / 284B Flash), so agencies with sustained workloads can serve them directly via vLLM-class tooling instead of paying API rates.

Sources: DeepSeek API pricing (live 2026-08-16) · DeepSeek release note — news260813 (price change 16:00 UTC Aug 16) · Anthropic — Claude Fable 5 pricing ($1.00 cache read) · @JulianGoldieSEO — the 276x cache-read gap (Aug 16, 2026) · Julian Goldie SEO — DeepSeek V4 Pro vs Fable 5 vs Grok 4.6 (full post) · TNW — DeepSeek warns of a 'significant' price rise (Aug 6, 2026, background) · SCMP (Aug 6, 2026, background)

Open-weight models are now a real alternative to paid frontier APIs. Alibaba's Qwen 3.8 Max (GA Aug 2–3, 2026; 2.4T-parameter MoE, ~95B active, 1M-token context) prices at $2 per 1M input tokens and $6 per 1M output tokens — the cheapest open frontier-class API on this page — with open weights live on Hugging Face since Aug 12, 2026 (Qwen/Qwen3.8-2.4T-A95B, plus an FP8 variant; official Qwen3.8-Max License — see the finalized-terms alert above). Moonshot's Kimi K3 — a 2.8T-parameter open-weight mixture-of-experts model (~104B active, 1M-token context, weights live on Hugging Face since July 27, 2026) — prices at $3 per 1M input tokens and $15 per 1M output tokens, a fraction of flagship paid APIs, while scoring within a few points of Claude Fable 5.1 and GPT-5.6 Sol on vendor-run coding benchmarks. Zhipu's GLM-5.2 (open weights, MIT license, 1M-token context) is the strongest open-source coding model on Terminal-Bench 2.1. NEW Aug 26, 2026: the anonymous Ox Alpha model that topped OpenRouter usage was confirmed as GLM-5.3-Flash — 320B/18B MoE, natively multimodal, 1M context, MIT weights live on Hugging Face, and official API pricing at $0.15/$0.50 per 1M tokens (50% promo $0.075/$0.25 through Sep 9, 2026) — roughly one-tenth of the flagship GLM-5.3 rate and below DeepSeek V4 Pro off-peak on both axes. NEW Aug 28, 2026: Tencent Hy4 preview — 770B total / 49B active MoE, native 1M-token context, Apache 2.0 weights live on Hugging Face — lists at $0.834/$2.501 per 1M tokens on OpenRouter (cached input $0.042), roughly 40% below GLM-5.3 on input and 43% on output, and ~6x below Kimi K3 on output, though still above DeepSeek V4 Pro off-peak ($0.66/$1.98); Tencent's internal blind eval scored it 2.99/4 vs GLM-5.3 2.92 / Kimi K3 2.94 (tier parity, not a decisive win), and the preview status + ~770GB FP8 weights (not self-hostable on typical infra) are the caveats. NEW Aug 28, 2026: Qwen3.8-Flash — Qwen's new open-weight MoE (125B total / ~6B active, multimodal, ~95% sparsity, 262K native context 1M via YaRN, qwen-community-1.0 license) — lists at $0.15/$0.47 per 1M tokens (cache read $0.016) on QwenCloud and OpenRouter, below GLM-5.3-Flash at list on output and cache and below DeepSeek V4 Pro off-peak on both axes, with one of the lowest cache-read rates on this page; benchmarks are vendor-reported so far and it is a "Next" preview of the Qwen4 architecture. And now the cheapest option of all is not an API: Meta's Muse Glimmer (Aug 10, 2026; 30B dense, Apache 2.0, 4-bit quant under 20 GB) self-hosts on a single consumer GPU at electricity-cost marginal inference — see the Local vs API cost scenario for the break-even framing.

Real-world agency usage so far: Qwen 3.8 Max's headline marketing claim — "autonomous coding over 10+ days" — is an official claim, not yet independently replicated, and its Fable 5-beating ranking has been disputed by independent benchmark testing. The most-cited hands-on test so far (Aug 6, 2026, an agency-community builder with ~172K followers) reports building 45 real projects while ignoring benchmarks — 3D racing games, RPGs, websites, a full OS, a promo video, and autonomous workflows — with mixed results: some demos were poor, several builds looked better than Fable 5, and every project took only a few hours. The same tester's follow-up explicitly walked back the "destroyed Fable 5" framing: Qwen 3.8 Max shone on fast, multimodal, image-guided builds (including a single-file premium landing page), while Claude Fable 5 stayed stronger on huge, long-running projects that need consistency across massive contexts. Treat these as first-person, self-reported results with no linked artifacts or independent replication — useful as a delivery-speed datapoint, not as a benchmark.

Cost variability for open-weight models: published API list prices (Qwen 3.8 Max $2/$6, Kimi K3 $3/$15 per 1M tokens) are real, but total cost depends heavily on how you run the model. Qwen 3.8 Max's weights are now downloadable (live on Hugging Face since Aug 12, 2026), so self-hosting cost varies with hardware (DGX Spark-class vs cloud GPUs), quantization (MXPF4 vs full precision), context length, and utilization — plus the license layer: the Qwen3.8-Max License triggers a separate Qwen license only above $50M/12mo aggregate revenue for MaaS or AI Work Assistant businesses (see the alert above). Agencies that self-host trade a variable per-token bill for fixed hardware cost — the crossover point depends on your monthly token volume. Until you benchmark your own workloads, treat self-hosted cost as a range, not a fixed number.

What this means for agencies: model strategy is now a pricing lever. The calculator's Model Strategy selector reflects it — open-weight stacks trim the compute line (and lift margins ~5 pts), a local Muse Glimmer self-host trims it further (the cheapest factor on this page, ~0.85×), and frontier-only stacks carry a premium. Keep workflows model-portable across at least two providers, benchmark on your own workloads (vendor tables are not your client's workload), and treat AI spend as a managed line item, not a fixed cost. Read the full analysis of what open-weight models mean for agency margins

Sources: Alibaba — Qwen 3.8 Max blog · QwenCloud — Qwen 3.8 Max pricing · HF model card — Qwen3.8-2.4T-A95B (open weights, live Aug 12, 2026) · official Qwen3.8-Max License · HF — Qwen3.8-2.4T-A95B-FP8 · 45-project field report (X, Aug 6 2026) · Moonshot — Kimi K3 blog · Kimi K3 API pricing · HF model card — moonshotai/Kimi-K3 · zai-org/GLM-5 · Reuters — Alibaba Qwen revenue-share plan (Aug 7, 2026) · Tencent — Introducing Hy4 preview (Aug 28, 2026) · HF model card — tencent/Hy4-preview (Apache 2.0) · OpenRouter — Tencent Hy4 preview pricing ($0.834/$2.501) · Tencent release announcement · Bloomberg — Tencent Hy4 coverage (Aug 28, 2026) · QwenCloud — Qwen3.8-Flash official model & pricing page (verified Aug 27, 2026 — $0.15/$0.47/$0.016 per 1M) · HF model card — Qwen/Qwen3.8-Flash-Next (125B total / 6B active, multimodal MoE, qwen-community-1.0; released Aug 26, 2026) · OpenRouter — qwen/qwen3.8-flash ($0.15/$0.47) · OrcaRouter — Qwen 3.8 Flash release (sparsity/architecture) · Artificial Analysis — Qwen3.8-Flash-Next (medians $0.30/$1.20) · Unite.AI — Qwen3.8-Flash-Next previews Qwen4 architecture (Aug 26, 2026) · Intelligent Living — Qwen3.8-Flash matches DeepSeek V4 Pro · SGLang docs — Qwen3.8-Flash-Next cookbook · HN thread — Qwen3.8-Flash-Next (Aug 26, 2026)

Local vs API: Meta Muse Glimmer Cost Scenario

The cheapest-model answer is no longer just "which API." Meta released Muse Glimmer on August 10, 2026 (Meta Superintelligence Lab; distilled from the closed frontier model Muse Spark) — a 30B dense, Apache 2.0 open-weight model whose official 4-bit quant fits under 20 GB and runs on a single consumer GPU at >200 tok/s (RTX 5090: 233.4 tok/s with DFlash speculative decoding vs 74.9 baseline, a 3.1× speedup). For an agency running sustained agentic, eval, or coding volume, the marginal cost of inference collapses to electricity once the one-time GPU capex is amortized — a category that previously meant either a weak small model or expensive on-prem infra.

Muse Glimmer — verified specs	Value
Parameters	30B dense (~29.6B total: 28B text decoder + ~1.8B ViT-G/14 perception encoder)
License	Apache 2.0 — weights + all artifacts (BF16, 4-bit quants, DFlash drafter, encoder)
Context	131,072+ (NVIDIA: "120K+")
Quantized size	Under 20 GB — official 4-bit K-Quant-17GB GGUF file = 16,756,681,056 bytes (~15.6 GiB)
Hardware	Single consumer GPU — K-Quant-17GB targets 24 GB VRAM; K-Quant-Dynamic targets 32 GB
Speed	>200 tok/s on RTX 5090 (74.9 baseline 233.4 with DFlash); M4 Max 23.737.8; M5 Max 26.650.2; AMD AI Max+ ~24 tok/s; Radeon AI PRO R9700 ~53 tok/s
Local runtime (Ollama)	`ollama run muse-glimmer` — official 18 GB build / 128K context / text+image; Apple Silicon MLX build `muse-glimmer:30b-mlx` (21 GB) with DFlash + image input; `ollama launch` integrations include Claude Code, Codex, Pi, OpenCode, GitHub Copilot, OpenClaw, and Hermes Agent
Intended uses	Local agents, function calling, local coding, LLM-as-a-judge evaluation; multimodal (text+image in, text out); works with OpenClaw and Hermes Agent scaffolds

Local self-host vs hosted APIs — the cost scenario

For an agency running sustained agentic / eval / coding volume, the one-time capex of a 24 GB-class consumer GPU amortizes against per-token API spend. Above a token-volume break-even, local marginal cost ≈ electricity — undercutting every hosted API on this page (DeepSeek V4 Pro at $0.022/1M off-peak cache reads is among the lowest cache-read rates for agent runs — DeepSeek V4.1-Flash is lower at $0.003/1M off-peak; Qwen 3.8 Max $2/$6, Kimi K3 $3/$15, and premium frontier APIs are higher) — including the hosted version of Muse Glimmer itself: Meta has published no official API price, but Together AI hosts the same 30B at $0.35 / $1.50 per 1M tokens ($0.04 cached input), so a local box beats even the same model's API once your volume clears the hardware break-even.

Why local wins on volume: LLM-as-a-judge / evaluation is an explicitly named intended use — the highest-volume, lowest-margin-per-token workload agencies run. Offloading evals, batch coding, and background agents to a local Muse Glimmer box cuts the token bill without touching client-facing frontier calls. Privacy and data-residency become a feature: the model card and NVIDIA both cite private data processing and credential handling as design points, so security-conscious clients (HIPAA/SOC2-adjacent, NDA-bound) can get "AI that never leaves the building" at 30B capability.

Keep it honest — the caveats: K-Quant-17GB carries ~1.0% benchmark degradation vs full precision; full 131K context plus vision encoder plus DFlash drafter needs the 24–32 GB envelope, not a cheap 8 GB card; it is not frontier-grade on the hardest reasoning (HLE 22.0, GPQA 83.5 — roughly size-class parity); there is no audio modality; and local means you own uptime, ops, and security patching. The open-weight market also just got more competitive in the same week (NVIDIA's Nemotron 3.5 Lightning 30B MoE, Qwen3.6-27B, Gemma4-31B) — that is downward pressure on what "cheap" means broadly, both locally and in API pricing. So: "cheapest per workload," not "cheapest model."

Compare on the calculator: pick Local / self-hosted (Meta Muse Glimmer) in the Model Strategy selector for the cheapest compute factor per workload at sustained volume (0.85× vs open-weight 0.95×, with a +8 margin lift). Note the ranking changed on Sept 11, 2026: DeepSeek V4.1-Flash (0.84×) is now the cheapest per-token API factor, and local only wins once a 24 GB-class GPU is amortized — so verify against your own token volume before quoting a local-only stack.

Sources: Meta official model card — Muse-Glimmer-30B · Meta GGUF repo (K-Quant-17GB file listing) · Hugging Face blog — Muse Glimmer · NVIDIA blog — Local AI open source models & agents · TechStartups — Meta launches Muse Glimmer (Aug 10, 2026) · AMD blog — Run Muse Glimmer 30B on AMD · Ollama library — muse-glimmer (official 18GB / 128K build) · Together AI pricing — Muse Glimmer 30B $0.35/$1.50

Cache-Aware Agent Runs: DeepSeek V4 Pro vs Claude Fable 5.1

The number nobody's talking about is 276. On Aug 16, 2026, Julian Goldie pointed out that long-running AI agents re-read their instructions on every step — hundreds of times per task — so the price of a cache read dominates the bill, not the headline input price. DeepSeek's cache reads cost ~276x less than Claude Fable 5's on DeepSeek's launch pricing ($1.00 vs $0.003625 per 1M tokens). That 276x figure is now stale on BOTH sides: DeepSeek's official peak/off-peak increase landed the same day (16:00 UTC Aug 16), and on Sept 1, 2026 Anthropic cut Fable 5.1's cache reads 75% (to $0.25). The current cache-read gap is ~11x off-peak ($0.25 vs $0.022) and ~6x peak ($0.044). Even at the narrowed gap, cache-aware pricing remains the biggest single cost lever in this calculator — model it below.

The scenario: a long agent run that re-reads a fixed context (system prompt + instructions + accumulated work) before every step. At a 92% cache hit rate — Julian's stated assumption for agent workloads (directionally plausible because the same context is re-read every step, but not a verified OpenRouter statistic) — 92% of input tokens bill at the cache-read rate, only 8% at the cache-miss rate. Output tokens bill at the full output rate. The default inputs below model a real heavy-agent workload; change them to match your client's.

Context Re-Read per Step (tokens)

System prompt + instructions + accumulated context, re-sent every step.

Steps per Run

Agent actions per task — long runs re-read context hundreds of times.

Runs per Month

Agent tasks executed per month across clients.

Output Tokens per Step

Model output per step (actions, code, answers).

Cache Hit Rate (%)

92% is Julian Goldie's assumption for agent workloads — assumption, not verified.

DeepSeek Pricing Basis

Current off-peak: hit $0.022 / miss $0.66 / out $1.98. Current peak: hit $0.044 / miss $1.32 / out $3.96. Launch: hit $0.003625 / miss $0.435 / out $0.87.

Claude Fable 5.1 — Monthly Cost

—

$10/1M input · $50/1M output · $0.25/1M cache read

DeepSeek V4 Pro — Monthly Cost

—

Total Cost Gap

—

Fable 5.1 ÷ DeepSeek V4 Pro

Monthly Savings on DeepSeek

—

Same workload, cache-aware stack

How to read this

—

Cache-read price gap — the 276x explained

Cache-read price (per 1M tokens)	Price	Gap vs Fable 5.1 ($0.25)
Claude Fable 5.1 (Anthropic — official)	$0.25	1×
DeepSeek V4 Pro — launch pricing (≤ Aug 16, 2026 16:00 UTC)	$0.003625	~69× (vs Fable 5.1; Julian's 276× used Fable 5's $1.00)
DeepSeek V4 Pro — current off-peak (after increase)	$0.022	~11× (current gap vs Fable 5.1)
DeepSeek V4 Pro — current peak (after increase)	$0.044	~6× (vs Fable 5.1)

Sources: Anthropic official pricing (Fable 5.1 cache reads & refreshes $0.25/MTok) · DeepSeek official pricing (current peak/off-peak) · DeepSeek release note — price change effective 2026-08-16 16:00 UTC · @JulianGoldieSEO — the 276x cache-read gap (Aug 16, 2026) · Julian Goldie SEO — DeepSeek V4 Pro vs Fable 5 vs Grok 4.6 (full post)

Verification note: the 276× ratio is arithmetically correct on DeepSeek's launch pricing vs Fable 5's $1.00 ($1.00 ÷ $0.003625 = 275.9×) but is stale on both sides after the 2026-08-16 16:00 UTC DeepSeek increase and Fable 5.1's Sept 1 cache-read cut — current gap vs Fable 5.1 ~11× off-peak / ~6× peak (ABD Legacy verification, Sept 1, 2026). The 92% cache hit rate is an author claim, not a published OpenRouter statistic — treat it as an assumption and benchmark your own hit rate.

Model: monthly input = context-per-step × steps-per-run × runs-per-month; cache-hit input = input × hit-rate (bills at cache-read price), cache-miss input = input × (1 − hit-rate) (bills at input-miss price); monthly output = output-per-step × steps × runs (bills at output price). Fable 5.1: $10 miss / $50 out / $0.25 cache read (Anthropic, Sept 1, 2026). DeepSeek V4 Pro current off-peak: $0.66 miss / $1.98 out / $0.022 cache read; peak: $1.32 / $3.96 / $0.044. Launch basis: $0.435 / $0.87 / $0.003625. This is a per-token bill comparison — the calculator's Model Strategy selector (DeepSeek V4 Pro cache-aware vs Claude Fable 5.1) applies the same gap to retainer-level pricing.

DeepSeek V4.1-Flash: the 100× Cached-Input Spread

DeepSeek shipped DeepSeek-V4.1-Flash on Sept 10, 2026 under the API name deepseek-flash — a 552B-parameter MoE (8B active parameters at prefill, 16B at decode), native vision in the base model, a 1M-token context, MIT-licensed weights, and a price sheet whose off-peak cached-input rate is $0.003 per 1M tokens. That last figure is the story: on the same model, the peak-band uncached input rate is $0.30 — a 100× spread between two columns that most cost models flatten into one "input price."

The rows: the table below is the model's real price sheet — input, output, cached input (standard off-peak and peak), context window and modality — rendered by the calculator from its own price data, including the two legacy IDs that are now aliases and the row everyone else got wrong this week: DeepSeek V4 Pro is not being retired. DeepSeek announced on Sept 10 that deepseek-v4-pro traffic would route to V4.1-Flash from 04:00 UTC on Sept 14, and withdrew that plan on Sept 11 — its API docs now state V4 Pro service continues past Sept 14 with unchanged billing. Treat every "forced migration" deadline you read this week as retracted.

Model row	API name	Band	Input $/1M (cache miss)	Output $/1M	Cached input $/1M	Context · modality	Status
Rendering the DeepSeek price sheet…

—

Price this workload: the defaults below are the worked example — 10M input tokens, a 50% cache hit, 1M output tokens, off-peak. Change any input to match a client's agent run; the model rows above stay the source of truth for the rates.

Input tokens per month

Prompt tokens sent to the model, cached or not.

Cache hit rate (%)

Share of input tokens billed at the cached-input rate. Agents that re-read a fixed context every step run far higher than 50% — benchmark yours.

Output tokens per month

Completion tokens, billed at the full output rate on every band.

DeepSeek pricing band

Same model, same tokens — the band is a 2× multiplier on every rate.

GPT-6 Astra: The 272K Repricing Cliff

GPT-6 Astra (launched Sept 3, 2026; live in the API and ChatGPT Work/Codex Sept 4) has a context window of 1,050,000 tokens — and a pricing rule most calculators miss. OpenAI prices prompts with more than 272K input tokens at 2x the input and cache rates and 1.5x the output rate for the FULL request: $10 $20 per 1M input, $1 $2 cached input, and $50 $75 per 1M output. So an agent run that loads a long context does not pay a little more — it reprices every token in the request, which is why long-context Astra work is the most expensive stack on this page. Use the scenario below to see both bands: what a workload bills at standard ≤272K rates vs. what the same workload bills once any prompt crosses 272K input tokens. (The "$10/$50 doubles past 272K" shorthand is wrong on output — OpenAI's official rule is 1.5x there.)

Input tokens per request

Above 272,000 = the whole request reprices (2x input/cache, 1.5x output). Max input 922,000.

Output tokens per request

Reasoning tokens bill at output rates. Max output 128,000.

Requests per month

Agent tasks / API calls per month across clients.

Input cache hit rate (%)

Cached input bills at $1/1M (≤272K) or $2/1M (>272K); cache writes $12.50/$25.

AI Coding Agent Cost Transparency

AI coding agent pricing is not the per-seat number on the pricing page. That's the core finding of TrueFoundry's August 6, 2026 guide, and it's why this calculator treats model strategy, delivery risk, and retries as explicit cost levers rather than hidden line items. Coding agents bill through four structures — and the same team can see very different invoices under each:

Billing model	Examples	What agencies should know
Flat per-seat	Cursor Pro, Windsurf Pro, Claude Pro	Fixed monthly fee per developer — predictable, but overage charges can appear when limits are exceeded, and limits often aren't published clearly.
Seat + credits	GitHub Copilot Pro, Pro+, Max	Lower base fee plus a monthly credit pool. A frontier model can burn credits up to 8× faster than a standard model on the same task.
Pay-per-token API	Claude Code (API mode), OpenAI Codex	No per-seat charge; billing follows token consumption. Under sustained high-volume agent workloads it can become the most expensive option, and teams routinely misbudget it without usage visibility.
Subscription-bundled agent access	Grok Bot (SpaceXAI) — all SuperGrok, Cursor Pro and Cursor Teams plans (since Aug 26, 2026)	No per-token meter and no standalone price: always-on agents run 24/7 on their own cloud computer and are bundled into eligible subscription tiers. Beta, desktop (macOS) and iOS; enterprise waitlist. Zero surprise-bill risk by construction — the counterpoint to metered-billing blowups. Standalone economics not priced yet; authentication mechanics undisclosed.

Grok Bot (SpaceXAI) — agent option metadata: Monthly pricing: unknown / contact sales — no standalone price has been announced; in beta and included with every SuperGrok plan (from $30/mo), Cursor Pro ($20/mo) and all Cursor Teams plans (from $40/seat/mo) as of Aug 26, 2026; enterprise activation via the admin dashboard since Sept 3, 2026. Supported platforms: desktop (macOS app) and iOS. Always-on / cloud execution: yes — bots share a computer of their own in the cloud, so jobs keep running 24/7 even when the laptop is closed. Usage caps: not disclosed — no per-token meter and no published tier quotas as of Aug 12, 2026. Estimate assumption: because Grok Bot has no usage-metered price, it cannot be cost-modeled per-token; budget the eligible subscription cost as the line item (see AI Coding Agent Pricing in 2026 ).

Sources: SpaceXAI (x.ai) — Introducing Grok Bot (Aug 11, 2026) · MacRumors — Grok Bot Brings Always-On AI Agents to macOS and iOS (Aug 11, 2026) · Oflight — Grok Bot Explained

Six cost variables predict spend better than the headline price: usage volume (light autocomplete vs. all-day agent workflows can differ 20–50× in token consumption), model selection, billing model, team size, billing cycle (annual discounts often 15–20%), and plan tier fit. Free and entry tiers often carry quotas production teams exceed within weeks.

ROI note for agency pricing: if you resell or deliver AI-assisted work, model the raw-token line and the retry/failure line separately — "a task that costs one credit unit on a standard model might cost eight on a frontier model." This calculator's Model Strategy and Delivery Risk selectors exist exactly for that reason. Don't quote the sticker price; quote the consumption profile.

Source: TrueFoundry — AI Coding Agent Pricing: How to Choose the Right Plan (Aug 6, 2026). Full breakdown with the six-step budgeting checklist: AI Coding Agent Pricing in 2026 (findaiagency.com)

Consumer AI Agent Pricing: Meta Hatch vs ChatGPT & Claude

Meta is preparing to launch Hatch, its consumer AI agent that takes multi-step actions (shopping, email, calendar, booking) across apps like DoorDash, Etsy, Reddit, Yelp and Outlook, with launch targeted for late August or early September 2026 and a new model codenamed Watermelon targeted for October. The premium tier is priced up to $199.99/month — REPORTED, not official: The Information reports Meta has considered it, final pricing is not set, and Hatch has not shipped as of Aug 25, 2026. For agencies the $200 anchor matters because it is the consumer price point clients will compare your retainer against — not the same number as your per-agent infrastructure cost. Full explainer: Meta Hatch AI Agent Price: What Consumer AI Agents Cost

Provider	Product	Price	Billing	Availability	Model	Category	Source
Meta	Hatch (reported) Full breakdown	$199.99/mo — pending confirmation	Monthly	Launching in coming weeks (late Aug – early Sep 2026 target)	Watermelon (targeted October 2026)	Consumer AI agent	The Information via RuntimeWire
OpenAI	ChatGPT Plus	$20/mo	Monthly	Available now	GPT-5.6 Sol / Luna	Consumer AI assistant	openai.com
OpenAI	ChatGPT Pro	$100–$200/mo (5× / 20× usage tiers)	Monthly	Available now	GPT-5.6 Sol	Consumer AI assistant	openai.com
Anthropic	Claude Max	Up to $200/mo	Monthly	Available now	Claude (Fable 5.1-class)	Consumer AI assistant	anthropic.com

* Meta Hatch price is REPORTED, pending confirmation — The Information (Jun 4 + Aug 24, 2026, via PYMNTS and RuntimeWire) reports Meta has considered a premium tier up to $199.99/month for Hatch ("Hatch Plus", 5–10× daily capacity of the free tier). Final pricing is NOT set, and Hatch has not shipped as of Aug 25, 2026. ChatGPT / Claude prices are current list prices as of Aug 25, 2026.

Agency takeaway: a consumer subscription is not an agency infrastructure cost — model the workload, not the sticker price. See Meta Hatch AI Agent Price: What Consumer AI Agents Cost for the full analysis, or run the AI Agent API Cost Calculator for per-agent workload math.

Agent Plugins: Build Once, Run Anywhere

On August 6, 2026, an open, vendor-neutral specification called Agent Plugins 1.0.0 was published for packaging Agent Skills and MCP servers into portable plugins — announced by OpenAI's developer account alongside AWS, Cursor, GitHub, and Vercel. A plugin is a directory: a minimal plugin.json manifest, a skills/ folder, an optional mcp.json, and reverse-domain client-extension namespaces. Six clients support the format at launch: Codex, ChatGPT, Cursor, GitHub Copilot, Kiro (AWS), and VS Code — with AWS's Agent Toolkit compatible (30+ curated skills) and Google's Agents CLI / Data Agent Kit already shipping it. The Technical Steering Committee comprises Amazon, Cursor, Microsoft, OpenAI, and Vercel; Google is joining as a Core Maintainer. The spec is v1.0.0 but labeled a Working Draft, and it deliberately leaves out installation, distribution, permissions, sandboxing, and trust/provenance — all flagged for future versions.

What changes for your pricing: the old cost model assumed a fixed per-platform adaptation cost — the same skill repackaged, re-configured, and re-maintained for every client platform. With one portable plugin, that per-client reimplementation cost collapses. The calculator's Portability / Agent Plugins selector models the shift three ways:

Build-once amortization (setup discount). One package serves 2–3 compatible clients at ~15% off setup, or 4+ clients / resellable at ~30% off — instead of charging full adaptation per platform. The brief's suggested range is a ~60–90% discount off the adaptation slice per additional client; these factors apply the discount to the whole setup midpoint conservatively.
Shared maintenance stream (retainer factor). One package + optional per-client extensions replaces N drifting forks, so retainers scale slightly down (0.97× for 2–3 clients, 0.94× for 4+).
New revenue line (plugin packaging & distribution). Packaging the deliverable as a branded plugin bundle — and distributing it across the compatible-client market — is a priced deliverable, shown in its own result card ($750 for a 2–3 client bundle, $1,500 for a resellable plugin). Resellable plugin lines also lift margins ~3 pts.

One caution before you ship plugins: Agent Plugins v1.0.0 has no permission model, no sandboxing, no provenance verification, and no secrets handling. Agencies selling plugins should self-impose signing, code review, and least-privilege practices — and agencies buying plugins must vet them. That trust layer is itself a sellable compliance service, and the calculator's quote notes it whenever a plugin option is selected.

Client portability cuts both ways: lower switching costs are great for client trust, but they weaken "sticky" platform-based retainers — lock-in risk moves from the package format to the marketplace/install layer. Price the build once, but keep the retainer tied to ongoing value delivered, not to platform lock-in.

Sources: Google Developers Blog — Agent Plugins (Aug 6, 2026) · @OpenAIDevs announcement (X) · Agent Plugins Specification 1.0.0 · Compatible Clients · AWS Open Source Blog · Vercel blog · Read the full agency guide on Find AI Agency

Cloudflare Wallets & Agent Payment Rails: Budget Ceilings for AI Spend

On August 4, 2026, Cloudflare announced Cloudflare Wallets and cloudflare.pay during Agents Week — the buyer side of agentic commerce, pairing with the Monetization Gateway (seller side, waitlist opened July 1, 2026). The service gives AI agents a human-readable wallet handle for paying APIs, MCP tools, content, and AI inference within limits set by the wallet's creator. Handle reservations opened on announcement day (Aug 4–5, 2026); full wallet access — onramping/offramping funds, issuing Virtual Wallets, live x402 purchases — arrives "in the coming months" with fees still undisclosed.

The two-wallet model

Account Wallets — owned by humans/organizations using Cloudflare. Hold stablecoins, can add/remove funds, and delegate spending authority down to Virtual Wallets.
Virtual Wallets — owned by AI agents, operate via API keys. An agent's maximum spend is capped by the limit set by the Account Wallet owner.
Identity is optional: a Cloudflare account gets a unique web address / handle (e.g. research.example.cloudflare.pay) that works as a stable ID when interacting with merchants — like DNS for agent identity, built on the agent's existing cryptographic keypair (Web Bot Auth).

Creator-set guardrails

Spending cap / allowance (periodic) — the dominant cost variable. Cloudflare's worked example: "$100 per week budget for AI inference" per employee, or a $10 exploration budget for a cheap-to-try agent.
Approved merchant allow-list — agents can only pay listed merchants.
Maximum transaction size — per-purchase ceiling.
Manual override required — agents that hit a limit request a human-approved escalation; they cannot approve it themselves. Planned anomaly controls notify admins of unusual spending (e.g. unexpectedly fast spending) for review.
Prompt-injection immunity: caps are enforced at the wallet's API layer, not inside the model's system prompt — so a cap cannot be overridden by content the agent reads. This is the "outside-the-model enforcement standard" and the reason wallet rails matter for agencies running fleets of agents.

Payments & economics

Payments run on x402 — an open protocol that attaches payment instructions to HTTP requests (server returns HTTP 402 "Payment Required" with a price manifest; agent pays and retries with proof of payment). Originally introduced by Coinbase in 2025; Cloudflare co-founded the x402 Foundation.
Micropayment economics: stablecoins (USDC on Coinbase's Base chain used as the primary example) settle in seconds at ~$0.0001 fees, making sub-cent per-call API billing viable vs credit-card interchange (1.5–3.5% + fixed fee).
Fees, supported stablecoins/blockchains, custody partner, and onramp provider are all undisclosed — the calculator's wallet fee input defaults to 0% for exactly this reason. Treat any fee figure as an assumption until Cloudflare publishes pricing.

What this changes for agency pricing

Cost forecasting shifts from prediction to cap-constrained planning. Worst-case spend per agent becomes a known number (the cap) — the 2026 cost-blowup record (prompt-caching misses, retry loops, effort scaling) shows the risk is tail spend, not average spend, and wallet caps cut the tail.
Client billing caps become programmable and auditable. Provision a Virtual Wallet per client project with an explicit cap, approved vendors, and transaction-size limits; over-limit requests route to a human for a documented override — an audit trail that doubles as a billing dispute shield.
Reserve handles now. Handles are free to reserve today (first-come identity, like early DNS names); pricing comes later. Agencies reselling AI inference should reserve client-facing handles and watch the pricing announcement.
Use the estimator above: enter agents × per-agent monthly spend, set a wallet fee if your rail has one, and set the creator-set cap to see capped vs uncapped spend — including the no-wallet-fee default and the exceeded-cap edge case.
Compare the two rails: Cloudflare agent wallets vs Stripe Link — what each rail costs an agency breaks down wallet caps, x402 economics and the September 2026 Muse checkout.

Sources: Cloudflare Blog — The programmable wallet for the agentic Internet (Aug 4, 2026) · Cloudflare press release (Aug 4, 2026) · Help Net Security (Aug 5, 2026) · crypto.news (Aug 5, 2026) · TechTimes (Aug 4, 2026) — verified via parent research brief t_6e67321b.

FAQ — AI Agency Pricing

Price AI projects with a number you can defend.

Your Pricing Is Ready

Build and run client automations with HighLevel

Recommended tools

GoHighLevel

Make

n8n

More free AI business tools

Find AI Agency

My Business AI Audit

Price AI projects with a number you can defend.

Your Pricing Is Ready

Build and run client automations with HighLevel

AI agency pricing updates, weekly

Recommended tools

GoHighLevel

Make

n8n

More free AI business tools

Find AI Agency

My Business AI Audit