AI Coding Tool Costs 2026: Per-Seat Plans, Metered Agents

Published September 15, 2026Updated September 15, 2026By ABD Legacy LLC
AI coding tool costs 2026 Cursor Projects pricing Per-seat vs usage-based Cost of parallel coding agents Codex vs Claude Code pricing
What changed — September 2026. On September 10, 2026 Cursor launched Projects in beta: one coordinator agent plans the work and delegates implementation to parallel subagents running on cloud computers, spinning up a local agent when something has to be tested on your own machine [1][2]. No seat price moved for that — the meter did. The agent layer is billed in tokens at API pricing, so a delivery team's bill now scales with agent concurrency rather than with seats alone [4][5]. Every figure on this page was re-fetched from the vendors' own pricing pages and documentation on September 15, 2026: nothing is carried over from an earlier pricing cycle, and every number a vendor does not publish is labelled as not disclosed rather than estimated.

What do AI coding tools cost in 2026?

The seat is no longer the bill. As of September 2026 the published plan prices are modest and comparable — Cursor Pro $20/mo, Pro+ $60/mo and Ultra $200/mo with Teams seats at $40 or $120 per user per month; Codex Free, Go $8, Plus $20 and Pro from $100 per month with Business at $20 per user; Claude Code on Max from $100/mo with Team seats from $20 per seat [5][15][16]. The difference sits on top of the seat. Cursor's cloud agents — the machines parallel subagents run on — are “charged at API pricing for the selected model” in tokens, with a spend limit required before first use, while Codex's over-limit mechanism is ChatGPT credits [4][5][15]. Cursor publishes no price at all for Projects, the multi-agent surface it launched in beta on September 10, 2026 [1][2].

Published prices, tool by tool (September 2026)

Plan prices below are quoted in the vendor's own units and only where the vendor publishes them. Three of the five rows are complete; Cognition's and Windsurf's pricing pages could not be read at all on 2026-09-15 (HTTP 404 and HTTP 429), so those rows are marked not publicly disclosed rather than filled with a third-party figure [17][18].

Published prices, tool by tool. Vendor pricing pages and docs, fetched September 15, 2026.
ToolSeats and plan prices (as published)How the agent layer is billed
Cursor
(incl. Projects beta)		 Pro $20/mo · Pro+ $60/mo · Ultra $200/mo (individual). Teams: Standard $40/user/mo, Premium $120/user/mo with 5x the Standard limits on Agent; Enterprise custom. All three individual tiers list Cloud agents as included [5][6]. Cloud agents are “charged at API pricing for the selected model” (tokens), spend limit required, on-demand usage past the included limit at the same rates, billed monthly [4][5]. Cursor Token Rate $0.25/M tokens on Teams/Enterprise [5]. Published individual guidance: $60–$100/mo total usage for daily Agent users, $200+/mo for multiple agents or automation [5].
OpenAI Codex Free $0 · Go $8/mo · Plus $20/mo · Pro from $100/mo (5x or 20x Plus limits); Business $20/user/mo [15]. An API-key deployment is billed at API pricing [15]. Over-limit usage extends with ChatGPT credits — a credits currency, not token on-demand [15].
Claude Code
(Anthropic)		 Max from $100/mo (5x or 20x Pro usage); Team standard seat $20/seat/mo billed annually, $25 billed monthly, Premium seat 5x more usage [16]. Anthropic publishes no fixed weekly message or token counts for Claude Code on any plan; the plan page sells usage as tier ratios, and API/consumption billing is separate [16].
Devin
(Cognition)		 Not publicly disclosed in this pass — Cognition's pricing page returned HTTP 404 on 2026-09-15; no figure is quoted here, and no price comparison is made [17]. No verified rate, so none is stated. The vendor describes Devin as an autonomous engineer that plans, writes, tests and ships production code on its own [17].
Windsurf Not publicly disclosed in this pass — windsurf.com/pricing returned HTTP 429 on 2026-09-15, and the docs and plans pages returned 404; no figure is quoted here [18]. No verified rate, so none is stated.

Parallel-agent support: what each vendor publishes

The cost comparison only makes sense next to what each tool actually ships for multi-agent work, because that is the layer the meter sits on. Cursor ships the coordination pattern as a default product feature; Anthropic ships the same idea as an experimental construct behind a flag [1][2][14].

Parallel-agent support, as each vendor publishes it (fetched September 15, 2026).
ToolCoordinator / subagent surfaceStatus
Cursor Projects: a coordinator agent that “doesn't write code itself but directs other agents that do”, creates a plan and sends agents to implement and test parts of it in parallel, running “as many in parallel as the work needs”. Cloud by default on the Project's own computer; the coordinator spins up a local agent when something must be tested on your machine [1][2]. Beta since September 10, 2026, rolling out to all users; no published GA date [1][2]. No published concurrency cap [1][2].
Claude Code Agent teams coordinate “multiple Claude Code instances working together as a team, with shared tasks, inter-agent messaging, and centralized management”; in-session subagents are a separate construct that “work within a single session” [13][14]. Agent teams are experimental and disabled by default, enabled with an environment flag [14].
OpenAI Codex No coordinator/subagent orchestration surface appeared in the Codex pricing page fetched for this comparison; the cloud-side features it lists are things like automatic code review and Slack integration [15]. Not captured in this pass — recorded as absent from the pages read, not as a claim that none exists [15].
Devin Framed as a delegated engineer rather than an orchestration surface: “Devin plans, writes, tests, and ships production code on its own” [17]. Unpriced in this pass (pricing page unreachable) [17].

The Cursor row: Projects moves the meter, not the seat price

Cursor's seat ladder is unchanged by Projects: Pro $20/mo, Pro+ $60/mo and Ultra $200/mo for individuals, or Teams Standard at $40 per user per month and Premium at $120 per user per month, where Premium adds 5x the Standard limits on Agent; Enterprise is custom-priced [5][6]. All three individual tiers list Cloud agents as included [6].

Projects itself is orchestration on top of that substrate, and Cursor does not price it separately. The string “Project” does not appear on Cursor's pricing page or in its cloud-agent documentation, so there is no per-project fee, allowance or credit pack to quote [4][5]. What is published is the metering rule underneath: “Cloud Agents are charged at API pricing for the selected model”, with a spend limit required before first use [4]. Past the included usage, on-demand usage continues “at the same rates”, billed monthly, and Cursor states that requests are “never downgraded in quality or speed” [5]. On Teams and Enterprise, a Cursor Token Rate of $0.25 per million tokens applies on top of model API pricing for included, on-demand and BYOK usage [5].

Cursor's only published cost guidance for agent-heavy use is an individual band: daily Agent users “typically $60–$100/mo total usage”, power users running multiple agents or automation “often $200+/mo total usage” [5]. It is published for individuals rather than teams, so read it as an illustration of the bill's shape, not as an agency total. The rate that actually drives the number is per-model — for example Composer 2.5 at $0.50 per 1M input and $2.50 per 1M output on the same pricing page [5].

The capability Projects orchestrates is older than Projects: subagents on their own isolated virtual machines shipped in Cursor's August 19, 2026 changelog entry, and cloud agents stopped requiring a connected SCM provider on August 27, 2026 [1][2][3]. That is why the arrival of a multi-agent surface did not move the price list — it changed which line of the invoice grows.

Why the comparison changed shape: from per-seat assistants to compute-metered agents

Every tool in the table above still prices a human seat. The differences appear past the seat, when the work is done by agents rather than by the person in it. Cursor's seat layer is per-human and its compute layer is per-token, and Projects increases the second without changing the first [1][4][5]. The practical consequence for a delivery team is blunt: the seat price no longer bounds the bill. Agent concurrency does.

That changes what an agency is budgeting for. Under per-seat assistant pricing, headcount times seat price was a defensible estimate of the tool line, and a fixed-fee retainer could be priced against it. Under compute-metered parallel agents, the same delivery can be run with two seats and a large number of agent runs, and the variable line is the tokens those runs consume. The two layers now need two budget lines and two owners: the seat line is procurement, the token line is delivery.

Worked as arithmetic rather than as a quote: at Cursor's published Composer 2.5 rates ($0.50 per 1M input tokens, $2.50 per 1M output), one agent run that reads 250,000 tokens and writes 40,000 costs $0.125 + $0.10 = $0.225. Run twenty of those in parallel on the same task shape and one cycle is $4.50, which is the direction the vendor's own multiple-agents band ($200+/mo) points [5]. This is arithmetic on a published rate — not a published price for Projects, for which no rate exists.

Two vendor decisions are worth knowing before you model it. First, the over-limit currency differs by tool: Cursor sells on-demand usage at the same API rates, Codex extends usage with ChatGPT credits, and Anthropic publishes no fixed weekly counts for Claude Code at all [5][15][16] — so a cross-tool comparison has to convert between a token meter, a credit pack and a tier ratio. Second, Cursor has not published how Projects draws on plan limits, so whether a Project consumes the same Agent/usage pool as a cloud agent is something you have to test on your own account rather than read off a pricing page [1][4][5].

The line items no vendor publishes

Related: the delivery pattern behind the meter

This page prices the agents; the pattern that produces the cost is written up separately. Multi-Agent Coding Orchestration for Agencies at findaiagency.com/multi-agent-coding-orchestration-agencies walks the workflow stage by stage — one coordinator agent that plans and delegates rather than writes code, subagents created and managed on the coordinator's behalf, the local-agent leg for tests that have to run on your own machine, and the review ramp — sourced to Cursor's own blog post and changelog entry of September 10, 2026 [1][2][24]. Read the two together: that page explains the workflow, this one tells you which line of the invoice it lands on.

Model Your Agent Token Line →
Open the AI Agency Pricing Calculator →

Frequently asked questions

What do AI coding tools cost in 2026?

The published seat prices as of September 2026 are: Cursor Pro $20/mo, Pro+ $60/mo and Ultra $200/mo, with Teams seats at $40 per user per month (Standard) or $120 (Premium, which adds 5x the Standard limits on Agent); OpenAI Codex Free $0, Go $8/mo, Plus $20/mo, Pro from $100/mo and Business at $20 per user per month; and Anthropic's Claude Code on Max from $100/mo, with Team standard seats at $20 per seat per month billed annually ($25 if billed monthly). Those are only the first line of the bill. Cursor's cloud agents — the machines parallel subagents run on — are charged at API pricing for the selected model in tokens, with a spend limit required before first use, and Codex's over-limit mechanism is ChatGPT credits. Cursor publishes no price at all for Projects itself.

How much does Cursor Projects cost?

Not publicly disclosed. There is no Projects line item on Cursor's pricing page or in its cloud-agent documentation, and the string "Project" does not appear on either page (checked September 15, 2026). What is published is the layer Projects runs on: cloud agents are charged at API pricing for the selected model, a spend limit is required before first use, and past the included usage on-demand usage continues at the same rates, billed monthly, with requests never downgraded in quality or speed. On Teams and Enterprise a Cursor Token Rate of $0.25 per million tokens applies on top of model API pricing. Treat any figure quoted to you for Projects as an estimate, not as a vendor price.

Is Cursor billed in compute-minutes or credits?

No. Cursor's own cloud-agent documentation bills the compute layer in tokens at API pricing for the selected model, requires a spend limit before first use, and continues past the included usage as on-demand usage at the same rates. Credits are Codex's over-limit mechanism, not Cursor's: OpenAI's Codex pricing page describes extending usage with ChatGPT credits. A calculator field labelled "compute minutes" or "Cursor credits" is a modelling choice by whoever built it, not a vendor quantity — Cursor publishes no compute-minute metric and no credits currency.

How much does a parallel-agent team add to the seat bill?

No vendor publishes a per-agent, per-run or per-team rate, so the honest answer is a method rather than a number. Take the vendor's published per-model token rates, multiply by the tokens one agent run actually consumes in your codebase, then multiply by how many runs you plan to have in flight: at Cursor's published Composer 2.5 rates ($0.50 per 1M input, $2.50 per 1M output) one run that reads 250,000 tokens and writes 40,000 costs $0.225, and twenty such runs in one cycle is $4.50. Cursor's only published guidance is an individual band — $60-$100/mo total usage for daily Agent users and $200+/mo for power users running multiple agents or automation — which is not a team total, and is the only public number that shows the shape of the bill.

What do these vendors not disclose about pricing?

Cursor: no price, allowance or credit for Projects; no statement of how Projects draws on plan limits; no concurrent-agent ceiling; no per-PR or local-agent price; and no general-availability date, so the beta has no published end. Anthropic: no fixed weekly message or token counts for Claude Code on any plan — only tier ratios. Cognition and Windsurf: their pricing pages were unreachable when this comparison was compiled (HTTP 404 and HTTP 429 respectively on September 15, 2026), so this page quotes no figure for either and makes no price claim about them. Where a vendor is silent, the table says so instead of estimating.

Related reading

Sources & Methodology

Every price on this page is quoted from a vendor page fetched on September 15, 2026, and every figure a vendor does not publish is labelled as not disclosed instead of being estimated. Cursor's own documentation carries no publication date, so those pages are cited as “fetched 2026-09-15” rather than given a date [4][5][7].

Accuracy note: the seat prices, the API-pricing billing rule for Cursor cloud agents, the required spend limit, the on-demand continuation at the same rates, the $0.25-per-million-token Cursor Token Rate and the $60–$100 / $200+ individual usage bands are all taken from Cursor's own pricing page and docs as fetched on September 15, 2026; the Codex tiers and the ChatGPT-credits mechanism from OpenAI's Codex pricing page; the Claude Code seat prices and tier ratios from Anthropic's pricing page. Cursor publishes no price, allowance or credit for Projects, so none is stated. Devin and Windsurf are listed as not publicly disclosed because their pricing pages were unreachable on 2026-09-15 (404 and 429) and no third-party figure is used as a substitute. The productivity figures Cursor published with Projects (30% more PRs for new users; six times as many for Projects-primary users) are vendor-internal measurements without a published method, carry a selection-effect caveat in independent coverage [12], and are deliberately not used as agency outcomes on this page. Dollar figures in the worked arithmetic are computed from published per-token rates and are labelled as arithmetic illustrations, not as published prices.