OpenAI Astra Cost Outlook 2026: What Agencies Should Budget

Astra is OpenAI's next major model — first officially named on August 1, 2026 — and it is not released yet. That creates a problem for agencies that price client work on token math: there is no model card, no API rate, and no availability to test against. The only cost signal is a counterfactual (a $2,000 estimate for a research run, not a price). This outlook separates what is confirmed from what is speculation, gives you a comparison table with shipped models, and gives you three scenarios for budgeting before OpenAI publishes real rates.

What we know about Astra

The confirmed facts are narrow, and they come from OpenAI itself: Astra exists and is OpenAI's next major model (named Aug 1, referenced Aug 18 and Aug 28). On August 7, OpenAI disclosed it could not rule out Astra meeting the "Critical" cybersecurity capability threshold under its Preparedness Framework — the highest classification, which no prior OpenAI model reached — and paused some Astra workloads until stricter safeguards were in place. On August 18, OpenAI announced a two-week reinforcement-learning pause on deployment-bound models and put its largest planned frontier RL run on hold, with roughly 20% inference-compute monitoring overhead. On August 28, OpenAI notified SpaceX that it will wind down the contract supplying OpenAI models to Cursor (proposed cutoff November 12, 2026) and will not provide future models — explicitly including Astra — under that contract.

What is not confirmed: the leaked "mozaik-alpha-fdm" checkpoint codename, the leaked game/website outputs, a release date (the Sept 3-10 window is an unverified rumor), whether it ships as "GPT-6," and all pricing. For the full leak explainer — what Astra is, what the demos allegedly showed, and the counter-claim that they may be Claude Opus 5 output — see what Astra is and what it means for agencies on Find AI Agency.

The $2,000 math run is not a price

The single most repeated "Astra cost" figure online is a misread. On August 1, OpenAI published Ten advances in mathematics and theoretical computer science and credited "an internal version of Astra, our next major model" with producing ten results, each with a Lean 4 certificate. Then came the line that gets quoted as a price: "The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates."

That is a token-cost estimate for the research runs, computed at GPT-5.6 Sol's API rates — not a price for Astra, not a per-month cost, not a promise about what OpenAI will charge. It is useful for exactly one thing: it tells you the research team burned enough tokens on ten hard math problems to equal $2,000 of Sol usage. It tells you nothing about Astra's per-token rate.

Why cost matters for agencies

Three forces make Astra a real budgeting question even before a price exists:

• Token burn on "Max effort." The leaked demos were reportedly generated zero-shot (some 1-shot) on a "Max effort" reasoning mode, and the poster claimed "on Max the model tends to think ALOT more than Sol." That claim is unverified — but if a next-gen reasoning model spends substantially more tokens per task than Sol, per-task cost rises even at identical rates. Modeling token volume, not just rate, is the whole game.
• Monitoring overhead. OpenAI's own August 18 post estimates monitoring overhead at roughly 20% of the inference compute being monitored for Astra-with-tools. OpenAI has said this overhead is not billed to customers, so its effect on your invoice is indirect — but it is a real constraint on how OpenAI can price the model and how much capacity it can sell.
• Price-war context. OpenAI cut GPT-5.6 Sol API pricing over 20% on August 21, 2026 — promo rates $4/$20 per 1M input/output, through at least November 21, 2026 — in a competitive cycle against Claude Opus 5. A flagship replacement launching into that environment has pressure to land at or below Sol's promo, not above it. See GPT-5.6 Sol API pricing for the full window and OpenAI inference monitoring overhead for the ~20% context.

Cost-comparison table: Astra vs the 2026 frontier

Shipped-model rows are sourced from our own pricing pages (GPT-5.6 Sol API pricing, Claude pricing) and OpenAI's pricing page as of August 31, 2026. The Astra row is a speculative placeholder.

How to read this table (speculative methodology)

The Astra row is a placeholder, not a price. We built it from the leak evidence (TestingCatalog's report of the mozaik-alpha-fdm outputs and "Max effort" token burn), the Kingy AI rumor-tracker caveats (leaks unverified, no model card, no catalog entry), the $2,000 counterfactual from OpenAI's Aug 1 post, and the Cursor bar from OpenAI's Aug 28 decision. Every shipped-model row sources from our own pricing pages and OpenAI's pricing page. Nothing in the Astra row should be quoted to a client as a rate. The moment OpenAI publishes a model card or API id, this row gets real numbers — see the update checklist below.

Scenario planning for Astra pricing

Since no price exists, agencies should budget three ways. Each scenario states its assumption explicitly, so when OpenAI ships pricing you know which one broke.

Whichever scenario you plan for, the contract language matters: keep model-swap clauses (client agrees the agency may switch underlying models as availability changes) and token-burn clauses (client pays for actual usage above a baseline). The Cursor shift is a live example of why model portability is a cost control: see Cursor's OpenAI model cost shift.

What to update when OpenAI ships pricing

When OpenAI publishes Astra's model card or API rates, this outlook needs a specific refresh. Checklist for the update:

1. Model card URL and API model id — replace the placeholder row with the real identifiers.
2. Input / output rates per 1M — list and any promo pricing, with effective dates.
3. Cache and batch rates — cached input and Batch API (-50%) tiers, matching how we quote Sol.
4. Availability surface — ChatGPT vs API vs Codex, and whether it reaches Cursor through any channel.
5. Benchmark links — SWE-bench Verified / agentic-coding evals once OpenAI publishes them; replace "none public."
6. Scenario check — mark which of A/B/C held, and re-price any client work that assumed a scenario.

OpenAI's official pricing page is the source to watch: developers.openai.com/api/docs/pricing.

Frequently asked questions

Sources / disclaimer

• OpenAI (Aug 1, 2026): Ten advances in mathematics and theoretical computer science — Astra named; ~$2,000 token estimate at Sol rates.
• OpenAI (Aug 18, 2026): Pacing model development in an era of cyber-critical capabilities — two-week RL pause; ~20% monitoring overhead.
• OpenAI (Aug 28, 2026): Our decision on Cursor following its acquisition by SpaceX — wind-down; Astra barred from the contract.
• TestingCatalog (Aug 29, 2026): First outputs from GPT-6 "Astra" model — mozaik-alpha-fdm leak; Max effort; unverified.
• Kingy AI (Aug 29, updated Aug 31): OpenAI Astra Rumor Tracker — leaks unverified; no model card; no catalog entry.
• Quartz (Aug 2026): Astra solves 10 math problems for ~$2,000 — the counterfactual coverage.
• OpenAI API pricing page (Aug 31, 2026): developers.openai.com/api/docs/pricing
• Tom's Guide (Aug 29, 2026): OpenAI is leaving Cursor in November — Cursor bar context.
• ABD Legacy internal: GPT-5.6 Sol API pricing, Claude pricing — shipped-model table rows.

Accuracy note: This page is speculative by design. As of August 31, 2026, OpenAI has published no Astra pricing, model card, API id, or release date; the leaked outputs are unverified and disputed. The Astra table row and scenarios are planning placeholders, not prices. Shipped-model prices are list/promo rates as of the dates cited and move frequently — re-verify before quoting. Budgeting analysis, not financial advice.