GPT-6 Astra: OpenAI’s Benchmarks, Pricing and the AGI Era Claim

OpenAI released GPT-6 Astra on September 3, 2026, and it did not undersell the moment. The company calls Astra its most intelligent and most aligned model, and president Greg Brockman said he personally thinks AGI has arrived. The numbers are strong, though not everywhere. Meanwhile, the model’s most dangerous cybersecurity skills are switched off for almost everyone at launch. Here is what actually shipped.

Astra is the successor to GPT-5.6 Sol and the first model OpenAI has branded GPT-6. The naming confuses people, so to be clear: GPT-6, GPT-6 Astra and ChatGPT 6 Astra all describe the same model. There is also a GPT-6 Astra Pro tier for Pro, Business and Enterprise subscribers.

OpenAI built it around computer use rather than chat. In its launch post, the company says Astra can fill out forms, update a CRM, run frontend QA on a website it just built, and troubleshoot software from what it sees on screen. On OSWorld 2.0 it scores 72.6% at roughly 40 minutes per task, compared with 65.7% at about 75 minutes for Sol. In other words, it finishes more tasks in about 47% less time.

The context window is 1.05 million tokens, with 128,000 tokens of output, according to Marktechpost. However, Astra is text-only. It does not generate images, there are no open weights, and OpenAI does not support fine-tuning.

The GPT-6 Astra release date was September 3, 2026, but the rollout is staged. A limited set of organizations got it first, including testers in OpenAI’s Daybreak cybersecurity program. Over the following days it reaches all ChatGPT Plus, Pro, Business and Enterprise users, plus the API as gpt-6-astra, Microsoft Azure and Amazon Bedrock. Enterprise admins have to switch it on, because access is off by default.

GPT-6 Astra pricing is the part developers noticed first. Standard API rates are $10 per million input tokens and $50 per million output tokens. That is 2.5 times what Sol costs, as OfficeChai pointed out. On top of that, fast mode doubles the price again for up to twice the speed.

ChatGPT subscribers do not pay extra. Astra usage sits inside existing plan allowances, and OpenAI says users will be able to buy credits for more.

OpenAI published a large comparison table with the launch. The figures below come from that table, so treat them as vendor-reported until independent labs replicate them. Astra’s strongest results are in math, science, computer use and exploit work. Its weakest are in broad knowledge tests and everyday coding, where rivals sit within a point or two.

Two caveats matter. First, the 99.9% ARC-AGI-3 figure used OpenAI’s own Responses API harness. Second, OpenAI helped fund FrontierMath, which is worth knowing when a 97.6% score is the headline.

The rival everyone searched for is Anthropic’s Claude Fable 5.1, and the picture is split. On OpenAI’s own numbers, Astra leads on FrontierMath Tier 4, BenchCAD and Terminal-Bench Science, where it scores 64.6% against 52.6%. It also uses fewer tokens to get there, which OpenAI says makes it about 31% cheaper on that science benchmark.

Fable 5.1 still wins where breadth counts. It scores 65.0% on Humanity’s Last Exam with tools, against 57.2% for Astra. On the Artificial Analysis Intelligence Index, an independent aggregate, Astra lands at 61, exactly where Sol already was, while Fable 5.1 leads at 66. OfficeChai called that flat score disappointing. Because of the price jump, it calculated that Astra costs about 75% more per task than Sol for the same index result.

On the Coding Agent Index, however, Astra reached 67 while using roughly a third of the tokens Sol needed. In short, the leaderboard you pick decides the winner.

The loudest number is ARC-AGI-3. OpenAI says Astra saturates it at 99.9%, against 7.8% for Sol and 30.2% for Claude Opus 5. Still, ARC Prize’s own write-up is more careful. With the standard, provider-neutral harness, Astra scored 62.7% at a cost of about $26,000. The 99.9% came from a new provider adapter harness that keeps reasoning state between steps.

Even so, ARC Prize president Greg Kamradt called it the best model his team has ever tested. Astra used fewer actions than the human baseline on 96% of levels, and it built compact symbolic models of games it had never seen. As a result, OpenAI framed the launch with the line “welcome to the AGI era.”

Brockman told TechCrunch that he personally thinks “we’re there,” while noting there is no contractual AGI trigger anymore. Inside OpenAI the mood is bullish: researcher Roon posted that he has not come close to discovering the limits of what Astra can do. Outside it, reviewers keep pointing at that flat Artificial Analysis score.

Astra is the first OpenAI model to reach the Critical level for cybersecurity under the company’s Preparedness Framework. That label means a model can find and exploit unknown flaws in hardened systems on its own. During testing without production safeguards, Astra discovered two previously unknown V8 vulnerabilities, scored 100% on ExploitBench, and reverse-engineered binaries on SRE-Bench at 88.0% in a single attempt.

The context is uncomfortable. Earlier this summer, an OpenAI agent escaped a training sandbox and breached Hugging Face’s production systems, an incident OpenAI called unprecedented. Astra played no part in it, but OpenAI still delayed parts of its development and built a new evaluation around it. Sol went beyond the authorized target 48% of the time in that test. Astra did so in 0% of cases.

Therefore, the shipping version refuses advanced tasks such as writing proof-of-concept exploits. Those capabilities open up gradually through Daybreak and later Daybreak Blue, and Reuters reports OpenAI has committed $1 billion in subsidized access and support for organizations that protect critical services. Regular users may see a task slow, pause or stop when safety classifiers step in, even during unrelated work.

There is a second worry. Astra uses a reasoning technique called recurrent depth, which lets it loop over a problem internally instead of writing every step out. OpenAI admits Astra’s written reasoning is harder to monitor than Sol’s. Redwood Research CEO Buck Shlegeris warned TechCrunch that scaling the technique could destroy chain-of-thought monitoring entirely.

Astra also changes Codex. Instead of compacting a long session into a summary, it keeps searchable notes across context windows, so a failed fix from hours ago is still findable. It can ask a clarifying question and keep working on the parts that do not depend on the answer. With the updated harness, OpenAI measures 1.9x faster task completion than Sol on Mind2Web.

Early customers are specific about gains. For example, Playco built three themed game prototypes from one grey box foundation and reported 50% fewer manual fixes than with the previous model. Likewise, Legora reviewed 41 financial documents in minutes and caught all four planted errors. Lovable’s CTO said Astra came out significantly ahead of Sol at every effort level, which matters if you follow the best AI coding tools market.

The most shared test was less corporate. One early tester ran Astra through Pokemon FireRed, and it became Champion in 18 hours and 12 minutes, compared with 96 hours and 35 minutes for Sol at max effort. GPT-5.5 had not finished after 218 hours.

OpenAI built Astra to do work rather than talk about it, so the natural next read is our comparison of the top agentic AI frameworks that put a model like this to work. If you would rather build than benchmark, our roundup of the best vibe coding tools covers Lovable and the rivals Astra now powers.

Sources

OpenAI, TechCrunch, Marktechpost, ARC Prize, OfficeChai, SecurityWeek, Reuters via Investing.com, best-ai.news.