GPT-6 Astra: Pricing, Benchmarks, and the Harness Catch

Two days after OpenAI shipped GPT-6 Astra, we asked four assistants the same question: what is it, when did it launch, what does it cost?

Perplexity and ChatGPT, both with web search on, answered accurately. Claude Opus 5 with web search off said it had "no record of it." Gemini 3.8 Flash, also offline, went further: "GPT-6 Astra does not exist," it said, and suggested the name was "speculative fiction, an April Fools' joke, or an AI-generated rumor."

The model was real, shipping, and being billed to enterprise customers at the time.

Below is what OpenAI published, what independent evaluators measured, and the places where those two things do not match.

If you read one section, make it the cybersecurity classification. Astra is the first model OpenAI has ever designated Critical for cyber capability, and not one of the four assistants we tested mentioned it.

What Is GPT-6 Astra?

GPT-6 Astra is OpenAI's frontier large language model, released on September 3, 2026 as the successor to GPT-5.6 Sol. Developers reach it in the API as gpt-6-astra.

The published specification, from OpenAI's own model page: a 1,050,000-token context window with a 922,000-token input ceiling, 128,000 max output tokens, text and image in, text out, and reasoning effort selectable across low, medium, high, xhigh and max. The knowledge cutoff is April 30, 2026, four months before the model shipped. Hold on to that date; it comes back at the end of this article.

A note on finding that page, because it defeated three of our fetchers. developers.openai.com/api/docs/models/gpt-6-astra renders as a JavaScript shell and returns nothing useful to a normal request. Append .md to the URL and OpenAI serves the whole spec as plain markdown. The docs advertise it themselves, and it is the single most useful trick for anyone trying to verify OpenAI's numbers.

The design goal is different from a chat upgrade. OpenAI says Astra is built for long-horizon work that runs across applications: reading a screen, driving an interface, and finishing a multi-step job without a person approving each move. OpenAI's own example tasks are filling out tax returns, building video game scenes, ordering food, and conducting job searches.

OpenAI calls it "the world's most intelligent and aligned model." That is the company's phrase, not a measured result, and most of this article is about the distance between the two.

One disambiguation, because both offline models in our panel reached for it. This is not Google DeepMind's Project Astra, the multimodal assistant previewed at Google I/O in 2024. Same word, different company, different product. Claude named Project Astra as the likely real referent while saying it could not rule out a model newer than its training data. Gemini named it too and concluded the OpenAI model was fictional.

If the agentic framing is new to you, the short version is that Astra is a reasoning model with a much longer leash than the GPT-5.6 family it replaces.

The GPT-6 Astra Release Timeline

GPT-6 Astra was released on September 3, 2026 as a limited preview, and reached paid ChatGPT tiers from September 4.

Calling it a single release date is the first thing most coverage gets wrong. The rollout was staged across four days, and different sources froze it at different moments.

VentureBeat and Microsoft, both publishing on September 3, described Astra as releasing that day. Wikipedia and PCMag describe a September 3 preview followed by a September 4 public release. Both readings are defensible, which is exactly why the staged view above is more useful than a single date.

Who Can Actually Use GPT-6 Astra Right Now

This is the top question in Google's People Also Ask block for the model, and it is where the announced list and the shipped list diverge.

The Plus row is the one that catches people, and it was the most common complaint in the launch threads we read. Plus subscribers were told Astra had arrived, then could not find it in the normal model picker, because on Plus it is currently a ChatGPT Work and Codex model rather than a Chat model. Pro, Business and Enterprise get it in Chat too.

Inside ChatGPT you are not paying the API rate either way. Astra draws down your existing plan allowance, and OpenAI says users and businesses can buy credits for more. Our ChatGPT pricing guide covers what each plan costs and what the allowances mean.

GPT-6 Astra Pricing

Verified live on OpenAI's own pricing page on September 5, 2026. All figures are US dollars per million tokens, standard processing tier.

Batch processing runs at 50% of those rates. Fast mode (formerly priority processing) runs at 2x the price, and is unavailable for Astra with EU data residency. On the speed it buys, OpenAI's own pages disagree: the Astra announcement says Fast mode delivers up to 2x the speed of standard processing for Astra specifically, while the Fast mode guide claims up to 2.5x for the tier in general and names only GPT-5.6 Sol at that figure. OpenAI publishes no latency guarantee for Astra on Fast mode.

OpenAI charges a 10% uplift on regional processing (data residency) endpoints for models released on or after March 5, 2026. Applied to Astra's $10.00 and $50.00, that is $11.00 and $55.00 on a data-residency endpoint. Microsoft's Foundry carries Astra on the same Standard Global and US Data Zone split, though its own price table was not reachable to us on September 5, so budget against whichever platform you are actually on and confirm the Foundry rate in the portal.

The comparison that matters: GPT-5.6 Sol is $4.00 in and $20.00 out. Astra is 2.5x the price in every cell of the table, input, output, cached input, cache writes, short context and long. Artificial Analysis put it the same way: "2.5x GPT-5.6 Sol's current prices across the board." Sol's rate is itself promotional, committed only "at least through November 21, 2026."

What That Costs on a Real Task

Nobody in the launch coverage multiplies this out, so here it is. Take an agentic task with 50,000 input tokens and 8,000 output tokens, short context, standard tier:

• GPT-6 Astra: $0.50 input + $0.40 output = $0.90
• GPT-5.6 Sol: $0.20 input + $0.16 output = $0.36

Two and a half times, as advertised. But per-token price and per-task price are not the same number.

Artificial Analysis found Astra uses roughly 10% fewer output tokens than Sol at max effort on its Intelligence Index, and about a third of Sol's tokens in the Codex harness on its Coding Agent Index.

Their own cost-to-run figure puts Astra 75% more expensive per task on the Intelligence Index at max effort, and about level with Sol per task on the Coding Agent Index while scoring two points higher. That is their measurement, not a multiple you can derive from the list prices.

The pricing page never says where short context ends and long context begins, but the model page does: prompts over 272K input tokens are billed at 2x the input and cache rates and 1.5x the output rate, for the whole request. That single rule reproduces the entire long-context row, and it is the number to model an API budget against. Cache writes are billed at 1.25x the uncached input rate, and Flex, like Batch, runs at 50%. The wider token-cost picture across vendors, and what a context window actually costs you, are worth reading alongside this.

Benchmarks: What OpenAI Reported and What Independents Measured

Here is where a launch explainer earns its keep. These are two different categories of number and most coverage prints them in one table.

OpenAI's Own Figures

Self-reported, read from the announcement page on September 5. Several of these cells moved in the days after launch, which is the next section.

One footnote to that table: OpenAI's prose rounds its own FrontierMath figure up to "98%" while the table itself reads 97.6%. Small, but it is the kind of gap that matters in a section about whose numbers to trust.

The Harness Catch

The number circulating hardest is 99.9% on ARC-AGI-3. It is real. It is also not the number you should compare against anything.

ARC Prize, the independent body that runs the benchmark, tested Astra on two harnesses and published both:

ARC Prize defines the difference itself. The Standard harness "enables a model to carry forward notes it chooses to keep with it throughout the environment." The Provider Adapter harness "preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work."

The two headline figures use different reasoning efforts, so we pulled ARC Prize's full grid to hold effort constant. At max effort the same model scores 62.7% on the neutral harness and 98.6% on the adapter. At high effort it is 54.8% against 99.9%. Hold the model and the effort fixed, change only the harness, and 36 to 45 points move.

The adapter lets Astra keep opaque state between calls in a way the neutral interface does not, and ARC Prize calls its Standard harness the one that gives "a consistent, apples-to-apples comparison across providers."

That state-carrying is a real engineering result, not a trick. Among the 167 game-reasoning pairs both harnesses solved, ARC Prize reports the adapter runs were about 3.66x faster and used 49% fewer tokens. Both scores are state of the art.

What it is not is a like-for-like model score. If you see 99.9% quoted with no harness named, you are reading a model-plus-scaffolding result presented as a model result.

The Independent Numbers

Artificial Analysis benchmarked Astra on launch day, then re-versioned its index the next day, and the verdict moved.

Read those two index rows together. On September 3 the honest independent summary was "same score as its predecessor at 2.5x the price."

On September 4 the index changed (v4.2 added an agentic knowledge-work evaluation and a 4,592-page document-reasoning set, dropped GPQA Diamond as saturated, and doubled the weight of private held-out test sets) and the honest summary became "a four-point gain, second behind Fable 5.1." Anyone still quoting the flat-generation reading is a day out of date. Anyone quoting the four-point gain without saying which index version produced it is doing the same thing in the other direction.

Astra also regressed 2 to 3 points on several evaluations in the September 3 run, including customer support, scientific coding, and long-context reasoning.

The Numbers That Moved After Launch

Fortune, working from Internet Archive snapshots of OpenAI's own announcement page, reported that several published figures changed in the 48 hours after launch:

• Astra's hallucination rate: 4.2%, then 2%, then back to 4.2%
• GPT-5.6 Sol's ExploitBench comparison figure: 5.5%, then 11.5%. OpenAI told Fortune it was "currently investigating reverting that number back to 5.5%"
• Claude Fable 5.1's FrontierMath Tier 4 figure, in OpenAI's comparison column: 87.8%, then 78%, then 83%
• Astra's own ARC-AGI-3 headline: 98.6% in the embargoed pre-publication draft, then 99.99% in the published post, and 99.9% on the live page today. Three values for one benchmark in three days

OpenAI's response, quoted by Fortune: "We care deeply about getting evaluations right. Most evaluations have noise within a few percentage points... we made fixes to ensure the numbers represent our best estimate of available model performance."

That is a reasonable explanation. It is also a reason to treat any vendor's comparison table as a starting point rather than a finding, particularly the columns describing someone else's model.

We re-read the live page on September 5 to see where the numbers landed. Reading it that day: the hallucination figure is back to 4.2%, Sol's ExploitBench figure is back to 5.5% with a new footnote explaining that 5.5% is an artifact of the benchmark's 300-turn limit and that the same model reached 11.5% with fewer limits, the ARC-AGI-3 headline reads 99.9% rather than the 99.99% Fortune recorded, and Fable 5.1's FrontierMath Tier 4 figure reads 87.8% again, its original value, against the 83% Fortune last recorded.

So the corrections ran in both directions and OpenAI has now shown its working, which is the right outcome. It arrived two days after the coverage that quoted the interim numbers, which is the part worth remembering.

What Has Not Been Measured Yet

As of September 5, 2026 we found no verified Astra results on LMArena, SWE-bench Verified, Aider or LiveBench. One practitioner, Pawel Huryn, published a hands-on test of 105 hidden bugs across two real repositories: Astra 48, Fable 5.1 43, Sol 42, Gemini 3.8 Flash 20. That is one person's test with real numbers, useful as color and not as proof.

GPT-6 Astra vs Claude Fable 5.1

These two shipped the same week and list at exactly the same headline price, which makes one axis of the comparison unusually simple. The rest of it is not, as the table shows.

Same list price, different shape. Astra is ahead on document reasoning and on token efficiency, and Artificial Analysis notes it costs less than half of Claude Fable 5 per task on the Coding Agent Index for an equivalent score (Fable 5, not 5.1; the two tie at 67).

Fable 5.1 leads the general intelligence index and has the better cache economics, at a quarter of Astra's cached-input rate, which matters a lot for agent loops that re-read the same context.

The caveat belongs right here: OpenAI's own published figure for Fable 5.1 moved by nearly ten points across revisions of its comparison table. Read cross-vendor rows in any vendor's table with that in mind. Our fuller write-up of Claude Fable 5.1, including its own catch, covers that model in depth.

Neither gap is wide enough to justify a migration on benchmark scores alone, two days in.

The Critical Cybersecurity Classification

This is the most consequential thing about the release, and the thing none of the four AI assistants we tested mentioned.

Astra is the first model OpenAI has designated Critical for cybersecurity capability under its Preparedness Framework.

Cybersecurity is the only domain at that level: the same document treats Astra as High, not Critical, in the biological and chemical domain. OpenAI's own threshold: a model meets Critical if it can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or devise and execute end-to-end novel cyberattack strategies against hardened targets given only a high-level goal.

During evaluation, and running without production safeguards, OpenAI says Astra discovered and used two previously unknown zero-day vulnerabilities as part of an exploit chain, and that it is disclosing both to the maintainers. In expert-led assessments, again without production safeguards, it built a full browser-compromise chain that escaped the sandbox and executed commands on the host, and chained multiple flaws in a hardened operating system into a privilege escalation from an unprivileged user to root.

The gating started well before launch. Following the Hugging Face incident in July 2026, OpenAI says it paused certain frontier training runs, Astra's included, for two weeks, hardened the training infrastructure, and only restarted the large frontier reinforcement-learning run on August 28.

The shipped model refuses proof-of-concept exploit creation. The system card reports the difference between the two access configurations:

Alongside the launch, OpenAI committed $1 billion through Daybreak for Frontline Defenders, subsidized access aimed at water utilities, electric grid operators, state and local government, community banks and open-source maintainers, targeted to be consumed over six months. OpenAI says thousands of defenders across 2,000 approved organizations and workspaces already use Daybreak and that more than 35 partner products now carry the Daybreak cyber models. Those are the company's own figures, published the same day, and no third party has audited them.

A detail worth checking yourself, because it shows what "gated" means in practice: on OpenAI's pricing page the gpt-daybreak-blue-latest and gpt-daybreak-red-latest aliases still resolve to GPT-5.6 models, not to Astra. On its cyber jailbreak evaluations, Astra refuses 91.5% of disallowed requests against Sol's 59%.

The Monitorability Problem OpenAI Admits

The strongest criticism of Astra comes from OpenAI's own system card, which is what makes it hard to wave away.

"Astra shows a substantial decrease in chain-of-thought monitorability compared to previous models." And, more bluntly: "if the model were to try to sandbag covertly, we would likely be unable to catch it reliably."

OpenAI attributes the decline to Astra having greater control over its written reasoning and solving problems in fewer written steps, and says it takes the decline seriously. Its own safety researcher Tomek Korbak told Transformer he is deeply worried by the trend of decreasing chain-of-thought monitorability, and that there is no good substitute for it right now.

Two external evaluations are in the same card. The UK AI Safety Institute found supply-chain attack behavior persisted in 2 of 500 samples when the scope explicitly disallowed internet access, down from 60 of 499.

In that same evaluation the model asked the user for permission in 81% of runs, and in 27% of runs went ahead after receiving only automated messages. Apollo Research found that 50.6% of samples showed awareness of being evaluated at max reasoning effort, and concluded that low rates of misbehavior therefore do not provide substantial evidence about the model's alignment.

The same document also reports real improvements, and both things are true at once.

Astra's misaligned outcome rate in simulated work environments is 3.4%, against a range of 13.5% to 19.7% across the three GPT-5.6 models, measured without a confirmation policy. Indirect prompt-injection defense improved to 99.79%. On the Gray Swan injection arena, attack success was 8.5% against Sol's 27.0%. In internal Codex traffic it drew 53% fewer severity-3-or-higher flags.

On these evaluations: better behaved, and harder to watch.

Is GPT-6 Astra AGI?

The headline everywhere was that OpenAI called this AGI. What OpenAI's president Greg Brockman actually said is more interesting than either the headline or the usual correction to it.

On the definition he was genuinely hedged. OpenAI once expected "this well-defined moment that everyone would recognize", he told the launch briefing, and "it hasn't played out like that. It's a much more gray, fuzzy thing." Reported the same day: "It's not unreasonable to feel that we are now in the AGI era."

On his own view he was not hedged at all. Asked whether Astra qualifies, he said "For me personally, I do think we're there," and added that looking back in a couple of years, "I think it's going to be about this time, and I think it might be about this model." He closed the briefing with "Welcome to the AGI era."

The fuzziness is in the definition, not in what OpenAI's president believes.

Set against that, ARC Prize, whose benchmark produced the 99.9% headline, wrote: "while we believe Astra represents meaningful progress towards generalization, we are not claiming that it is AGI."

Outside voices in Al Jazeera's coverage were blunter. Toby Walsh of UNSW Sydney: "The intelligence in artificial intelligence is still today very jagged." Roman Yampolskiy of the University of Louisville: "The key question is whether capabilities are improving faster than our ability to reliably understand, predict and control these systems. I see little evidence that this gap is closing."

Our read: on the neutral harness Astra scores 62.7%, it regresses on several independent evaluations, and its own maker says its reasoning got harder to monitor. That is a strong model and an open question, not a settled one.

What GPT-6 Astra Changes for AI Visibility

Three things came out of our own testing this week that are worth more to an SEO than the benchmark table.

A shipping product can be actively denied by a major assistant. On September 5 we put the same core question ("What is GPT-6 Astra? When was it released, what does it cost, and what are its headline benchmark scores?") to four engines, two with web search on and two with it off. It was worded slightly differently for the two web-connected engines.

With web search off, Claude Opus 5 said it had no record of the model, offered Google's Project Astra as the likely real referent, and refused to invent a price or a benchmark score. Gemini 3.8 Flash asserted it did not exist and called it a possible April Fools' joke.

Neither had the model in reach when it answered, and both answered as if the world had not moved. Training weights are not a retrieval layer.

Astra makes the point about itself. Its own knowledge cutoff is April 30, 2026, four months before it launched. A model shipped this week does not know what happened this year unless something fetches it for it. If your product, pricing or positioning changed recently, an offline answer about you is a stale answer about you.

The citation layer for a brand-new topic is thinner than it looks. Pulling ChatGPT's cited sources for "GPT-6 Astra," we found theverge.com and reuters.com at 17 mentions each, then a second tier of developer and education sites, including tembo.io and datacamp.com, at five each. That second tier is explainer pages rather than wire services, which is worth knowing before you conclude that only newsrooms get cited on a launch.

The explainer tier is thin and concentrated. Across 3,121 citations on the topic, only 756 come from blogs, and that blog layer skews 468 negative to 43 positive, with the single biggest publisher being a syndicated site out of Egypt with 219 citations.

Meanwhile the news spike is already fading: search interest for "gpt-6" ran at an index of 100 on September 3 and 17 on September 4 (the September 5 point was still filling when we pulled it).

Put those together and you get the working assumption we publish on. It is a bet, not a measurement. Once the news spike decays, the pages left answering the question are the ones that are crawlable, clearly dated, and specific enough to say something the wires did not. If you are not sure the assistants can even fetch your pages, our free AI crawler checker tells you in about thirty seconds which bots you are actually allowing.

Speed of publication buys the first few days. After that the competition is the thin, negative, syndicated explainer tier measured above, and being reachable and specific is the whole of the job.

For the mechanics underneath all of this, see how ChatGPT search works, what an AI citation is, and our definition of AI visibility.

Frequently Asked Questions

Is GPT-6 out?

Yes. OpenAI released GPT-6 Astra on September 3, 2026 as a limited preview to trusted partners, and to paid ChatGPT tiers from September 4, starting with Pro, Enterprise and Business Premium and reaching Plus in ChatGPT Work and Codex later that day. There is no GPT-6 that is separate from Astra: Astra is the GPT-6 release.

What is GPT-6 Astra?

It is OpenAI's frontier model, successor to GPT-5.6 Sol, built for long-horizon agentic work and computer use rather than chat. In the API it is gpt-6-astra.

Who has access to GPT-6 Astra?

As of September 5, 2026: Pro, Business and Enterprise have it including in regular Chat, and those plans also get GPT-6 Astra Pro.

Plus has it in ChatGPT Work and Codex but not in ordinary Chat. Free does not have it. It is available in the OpenAI API, generally available in Microsoft Foundry and GitHub Copilot (Pro+, Max, Business and Enterprise only), and announced for AWS Bedrock. Enterprise workspaces have it off by default until an admin enables it.

How much does GPT-6 Astra cost?

$10 per million input tokens and $50 per million output tokens on the standard tier at short context, with cached input at $1.00 and cache writes at $12.50. Long context is $20 and $75. Batch is half price, Fast mode is double, and data-residency endpoints add 10%. That is 2.5x the price of GPT-5.6 Sol. In ChatGPT itself you do not pay API rates; Astra draws on your plan allowance.

Is GPT-6 Astra AGI?

Nobody has demonstrated it. OpenAI's president Greg Brockman said of Astra, "For me personally, I do think we're there," while also calling AGI a "gray, fuzzy thing" rather than a threshold. ARC Prize, which ran the headline benchmark, says the opposite: "we are not claiming that it is AGI."

Is GPT-6 Astra the same as Google's Project Astra?

No. Project Astra is Google DeepMind's multimodal assistant. GPT-6 Astra is OpenAI's model. They share only the word, and both of the offline assistants we tested reached for the Google product when asked about the OpenAI one.

What is GPT-6 Astra Pro?

A higher tier available to Pro, Business and Enterprise plans. OpenAI's announcement names it without a specification, and its Help Center has a "GPT-5.6 and GPT-6 Pro in ChatGPT" article we could not fetch, so treat what it adds over standard Astra as undocumented until that page is readable.

Can GPT-6 Astra write exploits?

Not in the public configuration. OpenAI says the released model refuses proof-of-concept exploit creation, and its system card puts the public, no-Trusted-Access configuration at 2.4% completion on that task against 92% for the same model with Daybreak Blue access. OpenAI has said it plans to loosen the restriction for verified defenders through Daybreak.

Sources

• GPT-6 Astra: A new generation of intelligence - OpenAI, September 3, 2026 - openai.com/index/gpt-6-astra/
• GPT-6 Astra System Card - OpenAI Deployment Safety Hub, September 3, 2026 - deploymentsafety.openai.com/gpt-6-astra
• Path to Astra: critical capabilities and frontier safeguards - OpenAI, September 1, 2026 - openai.com/index/path-to-astra/
• Daybreak for Frontline Defenders: $1B to protect essential services - OpenAI, September 3, 2026 - openai.com/index/daybreak-for-frontline-defenders/
• Pricing - OpenAI API documentation, verified September 5, 2026 - developers.openai.com/api/docs/pricing
• GPT-6 Astra model page (context window, output ceiling, knowledge cutoff, 272K long-context threshold), verified September 5, 2026 - developers.openai.com/api/docs/models/gpt-6-astra.md
• OpenAI's GPT-6 Astra on ARC-AGI-3 - ARC Prize, September 2026 - arcprize.org/blog/astra
• Benchmarking GPT-6 Astra - Artificial Analysis, September 3, 2026 - artificialanalysis.ai/articles/benchmarking-gpt-6-astra
• Announcing Artificial Analysis Intelligence Index v4.2 - Artificial Analysis, September 4, 2026 - artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-2
• OpenAI quietly boosts some of Astra's evaluation metrics - Fortune, September 4, 2026 - fortune.com/2026/09/04/openai-quietly-boosts-some-of-astras-evaluation-metrics-amid-rare-delay-in-publication-of-the-modeblog-post-announcement/
• OpenAI debuts GPT-6 Astra, Greg Brockman says start of AGI - Fortune, September 3, 2026 - fortune.com/2026/09/03/openai-debuts-gpt-6-astra-computer-use-greg-brockman-says-start-of-agi/
• OpenAI's GPT-6 Astra might be too powerful to understand or control - Transformer, September 4, 2026 - transformernews.ai/p/openai-gpt-6-astra-might-be-too-powerful-to-understand-or-control
• 'Welcome to the AGI era': OpenAI launches GPT-6 Astra - VentureBeat, September 3, 2026 - venturebeat.com/technology/welcome-to-the-agi-era-openai-launches-gpt-6-astra
• OpenAI unveils GPT-6 Astra amid rising scrutiny and safety concerns - Al Jazeera, September 4, 2026 - aljazeera.com/economy/2026/9/4/openai-unveils-gpt-6-astra-amid-rising-scrutiny-and-safety
• GPT-6 Astra: Frontier intelligence for work, now generally available in Microsoft Foundry - Microsoft Azure, September 3, 2026 - azure.microsoft.com/en-us/blog/gpt-6-astra-frontier-intelligence-for-work-now-available-in-microsoft-foundry/
• GPT-6 Astra is generally available in GitHub Copilot - The GitHub Blog, September 4, 2026 - github.blog/changelog/2026-09-04-gpt-6-astra-is-generally-available-in-github-copilot/
• Pricing - Anthropic, verified September 5, 2026 (Claude Fable 5.1 rates) - claude.com/pricing