Update, September 25, 2026: This article is a record of GPT-6 Astra as released on September 3, 2026. On September 23, OpenAI added two models to the same GPT-6 generation: the balanced GPT-6 Sol and GPT-6 Luna for low-cost, high-volume work (our GPT-6 Sol and Luna guide). Astra remains the flagship of the GPT-6 generation. Claude Opus 5, which section 7 compares against, was followed on September 22 by Claude Opus 5.5, which is now the model Anthropic advises starting with for most workloads (Opus 5 remains available). The comparison table and price comparisons in section 7 reflect the time of writing.
Table of contents
- 1. Release overview: dates, specs, and who gets it
- 2. "I'm on Plus and it isn't there" — availability by plan
- 3. Pricing: the right way to ask whether $10/$50 is expensive
- 4. Benchmarks: the official table, and the traps in lining numbers up
- 5. The first model to reach "Critical" cyber capability
- 6. For developers: what breaks when you migrate
- 7. Compared with Claude Opus 5 and Gemini 3.8 Flash
- 8. Caveats: where the numbers come from, and what can change
- 9. Picking by use case: when Astra is the right call
- FAQ
On September 3, 2026, OpenAI released its new flagship, GPT-6 Astra (OpenAI API documentation). The API is generally available with no gating, the model ID is gpt-6-astra, and the context window is 1,050,000 tokens.
That said, "it launched" and "your plan can use it" turned out to be two different things. OpenAI itself announced a staged rollout starting with a limited set of organizations, and on ChatGPT Plus you cannot use it in regular chat (Chat). On Plus it lives in ChatGPT Work and Codex, and the official help pages still say so as of September 22, 2026. Section 2 stands on its own for readers who came here only for that one answer.
There is one more thing to keep in mind: benchmark values change depending on whose table you read. OpenAI's official announcement includes a score table that lines Astra up against Claude and Gemini — yet on the very same DeepSWE v1.1, Claude Opus 5 is 73.7% in OpenAI's table and 68.8% in Anthropic's own announcement. Section 4 prints the main values with their publisher attached and explains what to watch for when you put them side by side.
Updated September 22, 2026: We brought the availability-by-plan section in line with what the official help pages say now and added the benchmark values from OpenAI's official announcement to section 4. We also corrected the table in section 5: the 88.0% / 99.2% we had labeled "ExploitBench" are in fact SRE-Bench (binary reverse engineering) results. ExploitBench is 100.0% for Astra and 78.5% for GPT-5.6 Sol. On September 25, 2026, we added the Gemini 3.8 Flash column from the same OpenAI table to section 4 (73.8% on DeepSWE v1.1, within 1.4 points of Astra, Opus 5, and Sol).
GPT-6 Astra
Released September 3, 2026 / API generally available
Sources: OpenAI
1. Release overview: dates, specs, and who gets it
OpenAI positions GPT-6 Astra as its most intelligent model to date, describing it as state of the art at computer use, browsing, software engineering, science, and professional work. The official phrasing is that it excels at carrying out multi-step workflows that span code, the browser, and business software — a description that puts the emphasis on work that runs for a long time rather than single question-and-answer turns (OpenAI model guidance).
| Item | GPT-6 Astra | Notes |
|---|---|---|
| Release date | September 3, 2026 | Rolling out from a limited set of organizations |
| Model ID | gpt-6-astra | API is ungated |
| Context window | 1,050,000 tokens | — |
| Max output | 128,000 tokens | — |
| Knowledge cutoff | April 30, 2026 | Details in our list of cutoff dates |
| Reasoning effort | low / medium / high / xhigh / max | none is not supported |
| Supported features | Streaming, structured outputs, function calling, file search, image input, web search, prompt caching | — |
| Rate limits | 5,000–15,000 RPM / 1M–40M TPM | Depends on your usage tier (Build, Launch, Grow) |
Inputs above 272,000 tokens carry a surcharge. Input and caching double, and output goes up 1.5x. A million-token context window is priced so that the more of it you use, the higher your unit rate climbs, so it is worth treating "what fits" and "what pays off" as separate questions.
2. "I'm on Plus and it isn't there" — availability by plan
This was by far the most common complaint in the days after launch. The short answer: Plus does give you Astra, but not in the regular chat interface (Chat).
What follows is based on OpenAI's official help pages (GPT-5.6 and GPT-6 Pro in ChatGPT, ChatGPT Work and Codex, and Managing usage with GPT-6 Astra in Work and Codex) as they read on September 22, 2026. The key point: in Chat, Astra is offered under the name "GPT-6 Pro," with eligible plans and allowances that are separate from Work and Codex (the launch announcement calls it "GPT-6 Astra Pro").
Generally available. No gate, no waitlist — specify gpt-6-astra and it works. You pay for what you use.
In Work and Codex, the full allowance can go to Astra. GPT-6 Pro is also available in Chat (Pro $200: 200 messages a week; Pro $100 and Business Premium: 50 a week, shared with GPT-5.6 Sol Pro).
Available once an admin turns it on. Access was off by default at launch. GPT-6 Pro in Chat also depends on the workspace's model-access permissions.
Limited Astra usage in Work and Codex (credits for anything beyond that). GPT-6 Pro in Chat: 15 messages a month (shared with GPT-5.6 Sol Pro).
Work and Codex only, with limited usage (credits for anything beyond that). GPT-6 Pro in Chat is not included. Look in the model picker inside Work or Codex.
The launch announcement said Astra was rolling out "to a limited set of organizations" that day and would reach all Plus, Pro, Business, and Enterprise users, plus the API, Microsoft Azure, and AWS Bedrock, over the coming days (OpenAI's official announcement). The rollout ran late, and OpenAI's help pages explain that, to cover the delay, it handed out usage resets on September 3 and 4 and ran an automatic reset on September 7.
In short, people are looking in the wrong place. Plenty of users open ChatGPT, scan the model list, see nothing, and stop there — but on Plus the way in is Work and Codex.
✅ "Wait and it will reach regular chat on Plus" has now been ruled out by the official help pages. The initial announcement said all Plus users within days, and as of September 8, 2026, five days after launch, the picture was unclear enough that a thread asking for clarification was open on OpenAI's developer forum. The help pages now state plainly that "Plus includes Astra in Work and Codex, but not GPT-6 Pro in Chat," and add that having Astra in Work does not automatically include GPT-6 Pro in Chat. (Checked September 22, 2026. We will update this if it changes.)
3. Pricing: the right way to ask whether $10/$50 is expensive
The API price is $10 in / $50 out per million tokens. That is exactly double Claude Opus 5 at $5 / $25, so on unit price alone it sits at the expensive end.
In the same documentation, though, OpenAI argues that because it reaches better results with fewer output tokens, the estimated API cost per task is actually lower. What a reasoning model bills you is unit price multiplied by the tokens it actually emitted, so a 2x unit price still comes out cheaper if the output is less than half as long.
Do not take that at face value. "Cheaper because it writes less" is OpenAI's own estimate, not a third-party measurement. It also flips easily depending on the workload — fire off large volumes of short answers and the unit price hits you directly, and settings that make it think longer (xhigh / max) push output tokens up.
Run ten of your own representative tasks and look at the bill — there is no other way to know.
4. Benchmarks: the official table, and the traps in lining numbers up
OpenAI's official announcement includes a score table that sets Astra against GPT-5.6 Sol, Claude Fable 5.1 / Fable 5 / Opus 5, and Gemini 3.8 Flash. These are the main values.
| Metric (OpenAI's table) | GPT-6 Astra | GPT-5.6 Sol | Claude Opus 5 | Gemini 3.8 Flash |
|---|---|---|---|---|
| Terminal-Bench 4.0 (terminal work) | 57.9% | 37.3% | 52.6% | 19.1% |
| DeepSWE v1.1 (agentic coding) | 74.1% | 72.7% | 73.7% | 73.8% |
| Agents' Last Exam (professional work in real software) | 59.3% | 53.6% | 55.5% | — |
| OSWorld 2.0 (computer use, offline set) | 72.6% | 65.7% | 70.2% | — |
| AutomationBench (business automation) | 41.4% | 18.1% | 26.9% | — |
| ARC-AGI-3 (novel problem-solving) | 99.9% | 7.8% | 30.2% | — |
| FrontierMath Tier 4 v2 (math) | 97.6% | 83.0% | 73.2% | — |
| Humanity's Last Exam (with tools) | 57.2% | — | 63.6% | — |
Source: the table in OpenAI's official announcement (September 3, 2026). Per OpenAI's note, each value is the maximum across reasoning effort settings. A dash (—) means the table gives no value.
The biggest gaps are in terminal work, business automation, and novel problem-solving, which matches the framing in section 1 about work that runs for a long time. On the other hand, on DeepSWE v1.1 the top four sit within 1.4 points: Astra 74.1%, Gemini 3.8 Flash 73.8%, Opus 5 73.7%, and Sol 72.7%, and on Humanity's Last Exam (with tools) Opus 5 comes out ahead (with Fable 5.1 higher still in the same table). This is not a table where Astra wins every row.
Before you line them up: the same test can show different values in different tables
Claude Opus 5 on DeepSWE v1.1 is 73.7% in OpenAI's table and 68.8% in the table in Anthropic's own announcement (July 24, 2026). OpenAI's table carries no note on who measured the Claude value in that row, or with what settings. Both are published figures, but where the 4.9-point gap comes from cannot be determined from public information. On the DeepSWE run that third-party Artificial Analysis did for its own Coding Agent Index (run in Codex), Astra scored 68% and Sol 72% — the order is actually reversed (published September 9, 2026).
Astra's 74.1% is a DeepSWE v1.1 score. Setting it next to a Claude score from a different test, such as SWE-bench Verified, is not a comparison. The names may sound alike, but the tasks and the scoring are different.
On ARC-AGI-1, five models sit between 97.5% and 98.5%, and OpenAI itself writes that Astra saturates FrontierMath Tier 4, ARC-AGI-3, and ExploitBench. On metrics like these, a one- or two-point gap means nothing.
So what can you compare? Values inside the same table. Within OpenAI's table, that is Astra 74.1% vs. Opus 5 73.7%; within Anthropic's table, which we covered in our Opus 5 article, it is Opus 5 68.8% vs. GPT-5.6 Sol 72.7%. Pick numbers across tables and you can build either conclusion.
5. The first model to reach "Critical" cyber capability
This may be the bigger story of the two. OpenAI has classified GPT-6 Astra as the first model to reach Critical cybersecurity capability under its own Preparedness Framework (OpenAI Deployment Safety Hub, system card published September 3, 2026).
By OpenAI's own account, Astra can find previously unknown vulnerabilities and devise novel ways to exploit them against hardened systems without step-by-step human direction. In evaluations run with safety mitigations removed, it is documented as achieving a path to arbitrary code execution in a hardened browser and privilege escalation on a hardened operating system.
| Values published by OpenAI | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| ExploitBench (turning known vulnerabilities into working exploits) | 100.0% | 78.5% |
| SRE-Bench (reverse engineering binaries without source code, first attempt) | 88.0% | 55.9% |
| SRE-Bench (within four attempts) | 99.2% | 68.7% |
| Gray Swan IPI Arena attack success rate (lower is better defended) | 8.5% | 27.0% |
| HealthBench Professional (length-adjusted) | 63.4 | 60.5 |
| Severity-3-or-higher deviations in Codex simulations | 0.063% (34 out of 54,218) | 0.135% (73) |
Sources: OpenAI's official announcement and system card (September 3, 2026). ExploitBench was run without production safeguards.
The 100% on ExploitBench needs a discount. In the system card, OpenAI itself acknowledges that the results may be inflated because the model had been exposed to information about historical vulnerabilities. So it built an internal version (June–August 2026) using only vulnerabilities disclosed after Astra's knowledge cutoff, where Astra scored 39.0% and Sol 5.5% (OpenAI notes that Sol's figure reflects a 300-turn limit, and that Sol reached 11.5% at similar settings when it hit fewer limits).
Offensive capability went up, but as a target it has become harder to break. The success rate of indirect prompt injection fell from Sol's 27.0% to 8.5%. If you point agents at external web pages, that is an improvement you will feel in production.
OpenAI's response was not to dial the capability back but to draw a line by use case. It helps with defensive work such as secure code review and patching, while refusing offense-leaning requests such as writing proof-of-concept code for a known vulnerability. On top of that come broader monitoring of all tool-using reasoning, user confirmation for consequential actions, and adjusted safety boundaries for under-18s, with a separate vetted limited-access track for biology research and cybersecurity research.
That way of drawing the line runs through the other new models released in September too. Every vendor fenced off its most dangerous capabilities and opened the gate only to qualified parties. The era of shipping everything to everyone is ending — of the three back-to-back releases, that is the shift I think is most widely overlooked.
6. For developers: what breaks when you migrate
If you are switching over existing code, some of it will error out as written. These are the changes spelled out in OpenAI's model guidance.
Do not pass temperature, top_p, top_logprobs and the like. On reasoning models they not only do nothing, they can be rejected outright.
The model works with Chat Completions, but tool calling requires the Responses API. For agent work, treat it as mandatory.
none reasoning effortIf you were using none or minimal, OpenAI's instruction is to start from low and compare results.
If you are coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl (set to "30m"). Cache writes are billed at 1.25x the input rate, so compare cached tokens and total cost before and after the migration (OpenAI's migration steps).
On behavior, OpenAI explicitly notes that Astra asks the user clarifying questions more readily than before. That is welcome when a human is sitting there in a conversation, but it becomes a reason batch jobs stall when nobody is watching. If you want autonomous execution, say so in the prompt.
One more thing: Fast mode is unavailable in EU data residency environments, and it comes with no latency SLA. Keep that in mind if your design assumes a given response speed.
7. Compared with Claude Opus 5 and Gemini 3.8 Flash
Between September 1 and 3, 2026, Anthropic, Google, and OpenAI shipped new models on three consecutive days. Here is the current lineup, restricted to the items where published values can be confirmed.
| Model | Released | Input/output (per 1M tokens) | Cutoff | Context |
|---|---|---|---|---|
| GPT-6 Astra | September 3, 2026 | $10 / $50 | April 30, 2026 | 1,050,000 |
| Claude Opus 5 | July 24, 2026 | $5 / $25 | May 2026 | 1,000,000 |
| Claude Fable 5.1 | September 1, 2026 | $10 / $50 | June 2026 | 1,000,000 |
| Gemini 3.8 Flash | September 2, 2026 | $0.75 / $3.75 (introductory price, through December 31, 2026) | March 2026 | 1,000,000 |
Three things stand out.
① The top tier has converged on $10/$50. Astra and Fable 5.1 charge the same. Meanwhile Anthropic has not changed its guidance that most use cases should start with Opus 5, so the picture in which the $5/$25 Opus 5 is where the real contest happens continues.
② Fable 5.1 has the freshest knowledge. June 2026, followed by Opus 5 (May), Astra (April 30), and Gemini 3.8 Flash (March). That said, if you pair the model with web search, the gap barely matters in practice. Cutoffs bite when search is turned off.
③ Price is the only dimension separated by an order of magnitude. At its introductory rate, Gemini 3.8 Flash costs roughly one-thirteenth of Astra on input. Line the models up by intelligence and they bunch together; line them up by cost and they split by an order of magnitude — that is the shape of autumn 2026, and the gap is starker than it was at the time of our previous-generation comparison.
8. Caveats: where the numbers come from, and what can change
To be upfront about it: when this article first went up, it left out figures we could not verify. Those we have since confirmed on official pages have been added with their publisher attached.
At launch we could not confirm them on the publisher's page and left them out; we have since confirmed the table in OpenAI's official announcement and added it to section 4 (September 22, 2026). Some of the Claude values differ depending on who published them.
The official help pages now give estimates. For Astra in Work and Codex, per five hours: 5–45 messages on Plus and Business Standard, 25–225 on Pro (5x), and 100–900 on Pro (20x). These are estimated local messages, not fixed limits. The GPT-6 Pro allowances in Chat are listed in section 2.
Specs, pricing, and migration changes come from the API documentation and model guidance; benchmarks from OpenAI's official announcement (Claude values cross-checked against Anthropic's announcement); safety numbers from the system card; availability by plan from the official help pages. Every one of them was checked on the publisher's own page.
And one more thing. Information right after a launch moves. Availability changes on a scale of days, and pricing and limits can be revised. We re-checked availability by plan and the benchmark values against the official pages on September 22, 2026.
9. Picking by use case: when Astra is the right call
The work runs autonomously for a long time. Multi-step tasks that cross the browser and business software, terminal work, and agents that read external web pages (thanks to the improved resistance to prompt injection).
When you want to halve the unit price. On writing quality and everyday coding, nothing published so far shows a decisive gap. If Opus 5 covers you, it is the cheaper answer.
Price is what moves by orders of magnitude. For pushing summaries, classification, and drafts through in bulk, the introductory rate on Gemini 3.8 Flash is overwhelming. See also our overview of Gemini.
Get a feel for them before you pay. There is no harm in working through each vendor's free tier before deciding. Here is our comparison of the three free tiers.
Let me close with the most reliable way to decide. Rather than studying benchmark tables, it is quicker to run ten of your own representative tasks through two or three models with identical inputs. Time taken, amount billed, and the quality of what came back. Look at those three yourself and you will get a more accurate answer than any table can give you.
FAQ
Q1. Can I use GPT-6 Astra yet?
The API has been generally available since September 3, 2026, with no gate and no waitlist. In ChatGPT, Plus, Pro, and Business can use it in ChatGPT Work and Codex (with limited usage on Plus and Business Standard, and on Enterprise once an admin enables it). In regular chat (Chat), Astra is offered as "GPT-6 Pro" on Pro, Business, and Enterprise, and it is not included in Plus (official help pages as of September 22, 2026).
Q2. I'm on ChatGPT Plus and it isn't in the model list.
You are most likely looking in the wrong place. On Plus, Astra is available in ChatGPT Work and Codex, and it does not appear in the regular chat interface (Chat). Check the model picker there instead. OpenAI's initial announcement said all Plus users within days, but the official help pages as of September 22, 2026 state plainly that "Plus includes Astra in Work and Codex, but not GPT-6 Pro in Chat." If it is missing from Work or Codex in the desktop app as well, update to the latest version with Menu > Check for Updates.
Q3. Which is better, this or Claude Opus 5?
It depends on the use case. In OpenAI's official table, DeepSWE v1.1 is nearly a tie at 74.1% for Astra vs. 73.7% for Opus 5; Astra leads on terminal work (Terminal-Bench 4.0) at 57.9% vs. 52.6%, while Opus 5 leads on Humanity's Last Exam (with tools) at 63.6% vs. 57.2%. Note, however, that Anthropic's own announcement puts Opus 5 at 68.8% on DeepSWE v1.1, so the same test shows different values depending on the publisher. On price, Astra is $10/$50 and Opus 5 is $5/$25, so Opus 5 is the cheaper of the two.
Q4. It costs twice as much as Opus 5 — will it really run up a bigger bill?
That depends on the task. OpenAI says the estimated cost per task is actually lower because it emits far fewer output tokens, but that is the vendor's own estimate. Workloads that fire off large volumes of short answers feel the unit price directly, and making it think longer with xhigh or max increases output. Measure the bill on your own representative tasks.
Q5. What breaks when I migrate existing code?
Four things mainly. Drop the sampling parameters such as temperature and top_p; use the Responses API rather than Chat Completions for tool calling; note that the none reasoning effort is gone (OpenAI's instruction is to start from low and compare); and, if you are coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl. Cache writes are billed at 1.25x the input rate, so compare your bill before and after the migration.
Q6. Does "Critical cyber capability" mean the model is dangerous?
It refers to a capability level in OpenAI's Preparedness Framework, not a verdict on danger. Astra is the first model to reach that level, which is why OpenAI is drawing a line by use case. It helps with defensive work such as secure code review and patching, but refuses offense-leaning requests such as writing proof-of-concept code. If anything it is harder to attack than before: the success rate of indirect prompt injection dropped from GPT-5.6 Sol's 27.0% to 8.5%.
Q7. Is a million-token context window a better deal the more of it I use?
No. Inputs above 272,000 tokens carry a surcharge, doubling input and caching and raising output by 1.5x. In many situations, selecting the parts you actually need is cheaper than dropping in an enormous context wholesale.
Sources
Every figure in this article was confirmed on the publisher's page listed below. Nothing was copied over from roundup articles.
- OpenAI — GPT-6 Astra Model (API documentation)
Model ID, context window (1,050,000 / max output 128,000), pricing ($10 / $50 / cache $1 and $12.50), the surcharge above 272K tokens, the April 30, 2026 cutoff, reasoning effort levels, supported features, rate limits - OpenAI — Model guidance
The positioning language, the claim that estimated cost per task is lower because output tokens are fewer, the four migration changes (tool calling requires the Responses API; the cache setting is replaced), the Fast mode restrictions, the tendency to ask clarifying questions - OpenAI — official GPT-6 Astra announcement
Release date and the originally announced rollout, the benchmark table (DeepSWE v1.1, Terminal-Bench 4.0 and more, including the Claude and Gemini comparisons), ExploitBench (100.0% against Sol's 78.5%), SRE-Bench (88.0% on the first attempt and 99.2% within four, against Sol's 55.9% and 68.7%), and the internal ExploitBench (June–August 2026) at 39.0% and 5.5%. Table checked September 22, 2026 - OpenAI — Deployment Safety Hub: GPT-6 Astra
Critical cyber capability under the Preparedness Framework, the possible inflation of ExploitBench and the description of SRE-Bench (§10.1.2), Gray Swan IPI Arena (8.5% against 27.0%), HealthBench Professional (63.4 against 60.5), deviation rates (0.063% against 0.135%), the mitigations and the refusal policy. System card published September 3, 2026 - Anthropic — Introducing Claude Opus 5
DeepSWE v1.1 in the announcement's table (Opus 5 68.8%, GPT-5.6 Sol 72.7%). July 24, 2026 - Artificial Analysis — Benchmarking GPT-6 Astra
Third-party DeepSWE results from its Coding Agent Index, run in Codex (Astra 68%, GPT-5.6 Sol 72%). September 9, 2026 - OpenAI Help — GPT-5.6 and GPT-6 Pro in ChatGPT
GPT-6 Pro in Chat (powered by Astra) for Pro $100, Pro $200, Business, and Enterprise; Plus gets Work and Codex; Chat allowances (200 a week / 50 a week / 15 a month) - OpenAI Help — ChatGPT Work and Codex
Pro and Business Premium can use their full allowance for Astra; Plus and Business Standard get limited usage, with credits beyond that - OpenAI Help — Managing usage with GPT-6 Astra in Work and Codex
"Plus includes Astra in Work and Codex, but not GPT-6 Pro in Chat," the per-five-hour estimates, and the usage resets that covered the launch delay (September 3, 4, and 7) - OpenAI — GPT-6 Astra announcement (developer community)
The September 3, 2026 release date and the originally announced rollout - OpenAI developer forum — clarification thread on the scope of Plus access
The observation that access was announced for all Plus users while actual access is limited to Work and Codex (the situation as of September 8, 2026; since settled by the official help pages)
Related articles
- Claude Opus 5: the complete release guide — the current flagship it will be measured against
- GPT-5.6: the complete release guide — the previous generation and its Luna / Terra / Sol lineup
- Claude Fable 5.1: breaking changes and migration — the other new model released the same week
- Is GPT-6 Astra at low really cheaper? — cost and tokens measured effort by effort, against Sol
- Knowledge cutoff dates for the major AI models — how far each model's knowledge runs
