GPT-6 Astra landed on September 3, 2026, and the announcement page is doing several jobs at once. Some of what it describes is generally available in the API today. Some is capability-complete but gated behind a limited rollout. Some is behind an opt-in flag in a config file. And some is a name attached to a plan with no date, no criteria, and no definition.
Sorting those four categories is not a pedantic exercise. It is the difference between quoting a client for work you can do next week and quoting for work that depends on something OpenAI has not shipped.
This piece walks through what is actually available, what is gated, what carries a flag, what is announced only, the full price card including a surcharge almost nobody is reporting, and the things OpenAI published about its own model that cut against the launch.
The short version: the model is real and the plumbing changes around it are the most useful part of the release, but access to Astra itself is restricted, every computer-use number in the announcement is self-reported, and two of the named features in the post are one sentence long with nothing behind them.
What GPT-6 Astra actually is
The model ID is gpt-6-astra. It takes text and images in and produces text out. Context window is 1,050,000 tokens in and 128,000 out, with a knowledge cutoff of April 30, 2026.
Reasoning effort runs low, medium, high, xhigh, and max. Notably, effort none is not supported, and neither is a custom temperature or top_p, and neither are logprobs. Tool calling requires the Responses API. If you have code that sets sampling parameters, it will need changing.
The announcement’s own framing is computer use: "Anything you can do on a computer, Astra can do for you. Fast." Everything else in the release hangs off that claim.
The price card, and the cliff at 272,000 tokens
Per million tokens, standard processing:
| Input | Cached input | Cache write | Output | |
|---|---|---|---|---|
| Standard, up to 272K input | $10.00 | $1.00 | $12.50 | $50.00 |
| Standard, above 272K input | $20.00 | $2.00 | $25.00 | $75.00 |
| Batch | $5.00 | $0.50 | $6.25 | $25.00 |
| Fast mode | $20.00 | $2.00 | $25.00 | $100.00 |
The model it succeeds, GPT-5.6 Sol, is $4.00 in and $20.00 out. Astra is exactly 2.5 times the price of Sol on both input and output.
The row that is missing from most coverage is the second one. Requests carrying more than 272,000 input tokens are billed at double the input and cache rates and 1.5 times the output rate. If you are selling the million-token context window as a feature, that is the number to build the estimate on, not the headline rate. A whole-repository or whole-site-content pass crosses 272,000 tokens easily, and the moment it does, the effective price is $20 in and $75 out.
Flex processing is in beta at batch rates. Rate limits scale by tier from 500 requests and 500,000 tokens per minute at Tier 1 to 15,000 requests and 40 million tokens per minute at Tier 5.
Shipped and generally available: async tool calling and mid-turn steering
These are the two changes we would point a CMS developer at first, and they are getting almost no attention.
Async tool calling means the model keeps reasoning while your tool runs. OpenAI’s own description: "Async tool calling lets the model continue working after it calls a tool, without waiting for that tool’s result." Results are matched back later by call_id. Mid-turn steering lets you inject instructions over a WebSocket connection while a response is still in flight.
Both went generally available on September 3 and both require the Responses API. Both are supported on GPT-6 Astra and later models only. The constraints are worth knowing before you plan around them: they work with function and custom tools but not with hosted built-in tools, they are incompatible with programmatic tool calling, and they cannot be combined with parallel tool calls in multi-agent mode.
The reason this matters more than the benchmark deltas is latency shape. Any agent that calls something slow, a headless CMS build, an image pipeline, a payment provider, a slow third-party API, currently stalls while it waits. With async tool calling it does not. That changes how an agent product feels to use in a way no accuracy score captures.
There is no independent testing of either feature yet. Treat them as documented capabilities to evaluate, not as reviewed ones.
Shipped behind a flag: cross-context notes in Codex
This is the most substantively new engineering feature in the release, and it is opt-in. It applies to Codex rather than to the API.
Instead of compressing a long session into a single lossy summary when the context window fills, Astra keeps running notes across context windows, and earlier windows remain searchable. OpenAI describes it as being able to "keep notes across context windows, preserving accumulated details without repeatedly compressing them."
The status is unusually specific: "You can enable this experimental feature in your Codex config.toml, and it will become the default for Astra in the coming weeks." So it is shippable today, behind a flag, with a promise of a default and no date attached to it.
There is indirect external corroboration here, which is rare in this launch. ARC Prize, testing Astra on ARC-AGI-3, singled out the note-keeping behavior and described the notes as remarkable for their "precision and information density." That is a different workload from a code migration, but it is at least evidence from outside OpenAI that the mechanism does something.
For agency work, the use case is obvious: long refactors and framework migrations across a large codebase, where compaction currently loses the thread of why a previous fix failed.
Shipped, but you may not have it: computer use and the model itself
The computer-use tool in the API is generally available and lists gpt-6-astra as supported. The gate is the model, not the capability. ChatGPT’s release notes on September 3 are direct about it: "Access is rolling out to a limited set of organizations. Astra is not yet generally available."
Enterprise administrators enable Astra per workspace, and it is off by default at launch.
The reason for the restriction is in OpenAI’s own Preparedness Framework classification, which rates Astra Critical for cybersecurity and High for biological and chemical capability. OpenAI also warns that the extra safety checks may pause legitimate work, which is a caveat worth passing to any client planning to build on it.
Every computer-use number in the announcement is self-reported. OpenAI publishes 72.6% on OSWorld 2.0 against 65.7% for GPT-5.6 Sol, 92.7% on ScreenSpot-Pro against 76.9%, 41.4% on AutomationBench against 18.1%, and 59.3% on Agents’ Last Exam against 53.6%. As of this writing, none of those has been independently reproduced. Coverage that appears to corroborate them is, on inspection, citing OpenAI’s figures back. That is the ordinary condition of a launch-day benchmark table rather than anything unusual, and we set out how to read one in vendor benchmark scores are not leaderboard results.
Read 72.6% on OSWorld the way it should be read: roughly one task in four still fails. That is supervised assistance, not unattended automation, and it should be scoped and priced as such.
Announced only: Astra Pro and Private Safety Processing
Two names in the launch have nothing behind them yet.
GPT-6 Astra Pro. The entire mention is one sentence: "Users on the Pro, Business, and Enterprise plans will also get access to GPT-6 Astra Pro." There is no definition, no benchmark, no pricing, no separate model ID, and no explanation of how it differs from Astra. Treat it as a plan name rather than a documented product.
Private Safety Processing. Explicitly described as still in testing rather than shipped.
Neither of these should appear in a scope of work.
Named but not new: Daybreak was already shipping
Daybreak is the item in this announcement most likely to be misread, and reading the Astra page alone will get you there.
The page’s only mention is this: "Through OpenAI Daybreak, we plan to expand access and roll out less restrictive safeguards in the coming weeks." Read cold, that parses as a forthcoming program with no date attached.
It is not. Daybreak launched in August 2026 as a gated cyber-defense program, and OpenAI expanded it on September 3 with a $1 billion commitment to frontline defenders. It has a published tier structure: Blue for defensive work such as malware analysis, detection engineering, and patch validation, and Red for vulnerability research and controlled exploit validation, behind separate approval. It has an eligibility surface covering critical infrastructure, state and local government, nonprofits, and open-source maintainers. It carries a specialized model, GPT-5.6 Cyber, built from GPT-5.6 Sol.
It was also already running in someone else’s product. On September 3, the same day Astra launched, Cloudflare shipped context-aware vulnerability discovery and remediation in Managed Defense built on "OpenAI Daybreak models, including GPT-5.6 Cyber." That is available to selected customers by invitation during early access rather than generally, so it is not a GA claim, but it is a shipped third-party integration rather than a plan.
What is genuinely undated is much narrower than the program. The Astra page is not announcing Daybreak. It is saying that Astra’s less restrictive cyber safeguards will reach Daybreak customers at some unspecified point. Daybreak access itself has a mechanism and an application path today.
The general lesson is worth more than the specific correction: an announcement page is a poor source for the status of anything it mentions in passing, in either direction. Astra Pro reads like a product and is one sentence. Daybreak reads like a plan and has been shipping for a month.
Not a feature: Sites in ChatGPT, and "fast mode"
Two items in the announcement read as capabilities and are not.
Sites in ChatGPT is described as letting Astra "create, host, and share websites, web apps, and games directly from a prompt." Sites is a pre-existing ChatGPT feature. Astra is a new model running behind it. Nothing about Sites shipped with this launch.
Fast mode is a price multiplier, not a capability. It is described as delivering "up to 2x the speed of Standard processing" and it costs exactly twice as much. Worth knowing what that 2x is measured against: Artificial Analysis measured Astra at 64.3 output tokens per second with a 384-second time to first token at maximum effort, and described the model as slower than average. Fast mode doubles a slow baseline.
What OpenAI published against its own model
Two things in the launch materials are unusually candid, and both belong in any honest summary.
The first is in the system card, on chain-of-thought monitorability: "GPT-6 Astra shows a substantial decrease in chain-of-thought monitorability compared to previous models." The card goes further, saying that if similar degradations continue in future generations, OpenAI would "soon have significantly reduced confidence in detecting many forms of misaligned behaviors." That is OpenAI publishing a regression about its own flagship.
The second is the safety-check caveat already mentioned: the additional monitoring may interrupt work that is not doing anything wrong.
There is also a claim we would not repeat without hedging. The announcement says Astra "helped establish a stronger bound of 186" in work on large prime gaps, improving a term in a bound unchanged for more than eighty years. Proof pages are linked, but the announcement does not state that any of it has been independently verified or peer-reviewed, and "helped" is carrying a great deal of weight in that sentence.
What is actually being deprecated, and it is not because of Astra
No model is being retired because of this launch. There is no Astra-linked deprecation notice.
The migration deadline that should be on an agency’s calendar points at the previous generation instead. On December 11, 2026, gpt-5-2025-08-07, the GPT-5 mini and nano models, and o3-2025-04-16 shut down, with Sol, Terra, and Luna named as replacements. That was announced back in June. Separately, the legacy gpt-3.5-turbo-0125, gpt-4-0613, and gpt-4-turbo models shut down on October 23, 2026, and the Assistants API already shut down on August 26 in favor of Responses plus Conversations.
If you have client code still on GPT-5 or the Assistants API, that is the work to schedule. Astra is not.
What to do with this
If you are building agent features, the async tool calling and mid-turn steering work is available now and is worth an afternoon of evaluation regardless of whether you adopt Astra itself. If you run long Codex sessions, turn on cross-context notes and see whether it changes how much context you lose.
If you are quoting computer-use work, scope it as supervised. One in four OSWorld tasks fails on OpenAI’s own numbers, and nobody outside OpenAI has checked those numbers yet.
If you are budgeting, build the estimate on the 272,000-token tier if long context is part of the pitch, and remember that Astra is 2.5 times Sol on both sides of the meter. We looked at the same problem from the other direction in what a model actually costs to run: the sticker price is rarely the number that lands on the invoice.
And leave Astra Pro and Private Safety Processing out of anything you sign. Daybreak is real and has an application path, but check the tier you would actually qualify for rather than assuming Astra-level capability comes with it.
Frequently Asked Questions
When did GPT-6 Astra launch? September 3, 2026, alongside a system card and an independent write-up from ARC Prize published the same day.
Can I use GPT-6 Astra today? Possibly not. OpenAI’s ChatGPT release notes say access is rolling out to a limited set of organizations and that Astra is not yet generally available. Enterprise administrators must enable it per workspace, and it is off by default.
What does GPT-6 Astra cost? $10 per million input tokens and $50 per million output tokens at standard processing, with cached input at $1. Above 272,000 input tokens, input and cache rates double and output goes to $75. Batch is half the standard rate. Fast mode is double.
How does that compare to GPT-5.6 Sol? Astra is exactly 2.5 times Sol’s price on both input and output. Sol is $4 in and $20 out.
What is the context window? 1,050,000 tokens of input and 128,000 tokens of output. The surcharge above 272,000 input tokens applies to the whole request.
Is async tool calling generally available? Yes, as of September 3, 2026. It requires the Responses API and works with function and custom tools, not with hosted built-in tools. It cannot be combined with programmatic tool calling.
What are cross-context notes in Codex? A replacement for lossy compaction. Instead of summarizing a long session into one compressed block, the model keeps running notes across context windows and earlier windows stay searchable. It is opt-in through config.toml today, with OpenAI saying it will become the default for Astra in the coming weeks.
What is GPT-6 Astra Pro? Unknown. The announcement mentions it in a single sentence with no definition, benchmarks, pricing, or model ID.
Has anyone independently verified the computer-use benchmarks? Not as of this writing. OSWorld 2.0, ScreenSpot-Pro, AutomationBench, and Agents’ Last Exam figures in the announcement are all OpenAI’s own. Third-party results exist for other things, including ARC Prize’s evaluation and Artificial Analysis’s index, but not for computer use.
What is OpenAI Daybreak, and is it available? A gated cyber-defense program that launched in August 2026 and was expanded on September 3 with a $1 billion commitment to frontline defenders. It has two approval tiers, Blue for defensive work and Red for vulnerability research, and a specialized model called GPT-5.6 Cyber. It is available by application to eligible defenders now. What the Astra announcement leaves undated is only whether Astra’s relaxed safeguards will extend to it.
Does anything get deprecated because of Astra? No. The nearest live deadlines are unrelated: GPT-5 and o3 shut down on December 11, 2026, the legacy GPT-3.5 and GPT-4 models on October 23, 2026, and the Assistants API already shut down on August 26.
Does GPT-6 Astra support temperature and top_p? No. Custom sampling parameters are not supported, and neither are logprobs. Reasoning effort is set instead, from low through max, and effort none is not available.
What did OpenAI say about monitorability? That Astra shows a substantial decrease in chain-of-thought monitorability compared to previous models, and that continued degradation of that kind in future generations would significantly reduce confidence in detecting misaligned behavior. That statement is in OpenAI’s own system card.
