GPT-6 Astra: What OpenAI Shipped and What It Costs

OpenAI released GPT-6 Astra on 3 September 2026. The model has a 1,050,000 token context window, a 30 April 2026 knowledge cutoff, and list pricing of $10 per million input tokens and $50 per million output tokens. That is 2.5 times the price of GPT-5.6 Sol on both sides of the meter, and the launch post is not where you find that number. The pricing page is.

Update, 29 September 2026: OpenAI has cancelled the planned October release of GPT-6.1 Astra, the successor to this model, after safety tests. Our report on the cancellation covers what was said, and what to do when an AI coding agent says it is done covers the honest-reporting problem it raised.

If you are building on OpenAI models, the interesting parts of this release are in the API docs rather than the announcement. Here is what changes and what it costs.

What GPT-6 Astra actually is

Astra is a frontier reasoning model that accepts text and images and returns text. Per the model reference, it supports reasoning effort levels from low to max, streaming, structured outputs, function calling, file search, image input, web search, and prompt caching. It exposes 922,000 maximum input tokens and 128,000 maximum output tokens inside that 1,050,000 window.

OpenAI positions it as state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work, and reports scores of 98% on FrontierMath Tier 4, 99.9% on ARC-AGI 3, and 100% on ExploitBench. Those are the vendor's own numbers on the vendor's own selection of benchmarks, which is the normal starting point rather than the finish line. They are worth reading the way you would read any release note, sceptically and against your own workload.

Access is staged. Astra went first to a limited set of organisations in OpenAI's Trusted Access Program, with ChatGPT Plus, Pro, Business, and Enterprise plans plus the API, Microsoft Azure, and AWS Bedrock following over the days after launch.

What it actually costs

Two details in that table matter more than the headline.

First, the long context columns are not decoration. Astra charges 2x input and 1.5x output once a request crosses 272,000 tokens. Any agent that accumulates a long trajectory, a big repository dump, or a multi-hour session will cross that line routinely, and the effective rate for that workload is $20 in and $75 out, not $10 and $50. If you are budgeting from the headline number you will be wrong by roughly double.

Second, GPT-5.6 Sol's cheaper pricing is promotional and the pricing page commits to it only through 21 November 2026. The 2.5x gap you can measure today is a gap between one list price and one promotional price, and it can narrow without Astra changing at all.

Prompt caching softens both. Cached input runs at $1 per million against $10 standard, with cache writes at $12.50. For an agent that replays a large stable system prompt and codebase context on every turn, caching is the difference between a viable bill and an unpleasant one. The usual cost-control habits apply here unchanged, just with bigger numbers on both ends.

What is not in the box

Astra ships on Chat Completions and Batch, and through the Responses API with web search, file search, code interpreter, and computer use tools. It does not support Realtime, and it does not support fine-tuning. If your product depends on either, this release does not replace anything for you yet.

The knowledge cutoff is 30 April 2026, which is four months before launch. That is a normal gap and a normal trap. Anything the model asserts about the AI landscape since May 2026, including other models released this summer, is either retrieved at run time or guessed. The same caution that applies to every model training cutoff applies here.

The safety classification is the real headline

OpenAI's system card states that Astra meets the Critical cybersecurity threshold in its Preparedness Framework, which makes it the most capable model the company has broadly deployed under that classification, with a corresponding set of monitoring and access controls. That is the first time this has happened, and it is a bigger structural change than the price. We cover what the classification means, and what it changes for people who ship software, in a separate piece on the Critical threshold.

It also closes a loop from August, when OpenAI paused work on Astra over exactly this capability. The pause was not a cancellation. It was a gate, and the model came through it with extra controls attached.

What to do this week

Nothing urgent. Astra is a staged rollout, and a model you cannot get access to yet cannot be evaluated. When it does reach your account, the useful test is not a benchmark, it is your own traffic: run a slice of real requests through both Astra and whatever you use now, measure quality against your existing evaluation set, and put the token counts into a spreadsheet with the 272,000 threshold marked. A 2.5x price increase is worth paying when it removes a retry loop or a human review step, and worth skipping when it buys you a slightly better answer to a question you had already answered well enough. The general framework for that call is in our guide on telling whether a model release is a big deal, and on keeping up with AI news without letting every launch reset your roadmap.

For context on the competitive week: Anthropic made Claude Fable 5.1 and Mythos 5.1 generally available on 1 September, and Google shipped Gemini 3.8 Flash on 2 September. Three frontier vendors moved in four days.

An earlier incident in the same vein is covered in the OpenAI agent Medicare portal breach.