On September 1st, OpenAI's president Greg Brockman closed a briefing on the company's newest model with four words: "Welcome to the AGI era."
He stopped short of an official declaration. He just said he personally believes OpenAI built it.
Astra, publicly released as GPT-6 Astra, is the model behind that sentence. Before deciding what to make of the AGI talk, it's worth seeing what the model actually does, in its own numbers, next to the models it's supposed to have just outrun.
It gets more done, and gets there faster
Astra's biggest practical jump is in computer use, actually operating a computer the way a person would: filling in forms, updating a CRM, managing a calendar, researching online, troubleshooting software. On OSWorld 2.0, the benchmark built to measure exactly that, Astra scores 72.6 percent in about 40 minutes. Its predecessor, GPT-5.6 Sol, needed 75 minutes to reach 65.7 percent. Paired with updated infrastructure, OpenAI says the model now completes Mind2Web tasks 1.9 times faster than before.
It writes code the way a senior engineer reads it
On Terminal-Bench 4.0, a test of real command-line software engineering work, Astra scores 57.9 percent. Sol scored 37.3. Anthropic's Claude Fable 5.1 scores 55.8, close, but still behind. Google's Gemini 3.8 Flash manages 19.1, a wide gap. Astra also posts 74.1 percent on DeepSWE v1.1 and 64.5 on FrontierCode 1.1 Extended, and on a design-heavy benchmark called BenchCAD it hits 95.9 against Claude Fable 5.1's 84.3.
Jane Street, which builds trading systems for a living, said Astra "communicates in a way that's easier for developers to follow and produces code that requires less iteration to reach production quality." Cognition, the company behind the AI coding tool Devin, credited its "excellent computer use, writing, and codebase understanding."
It reasons at a level nothing before it has
The math and science numbers are the most startling of the set. On FrontierMath Tier 4, one of the hardest math benchmarks that exists, Astra scores 97.6 percent, up from Sol's 83.0. On ARC-AGI-3, a test built specifically to resist memorisation, it hits 99.9 percent. On GPQA Diamond, a graduate-level science exam, 96.0 percent. OpenAI says the model helped establish stronger mathematical bounds on a long-studied problem in prime number gaps: actual, if narrow, research output, not a demo.
The one skill OpenAI is nervous about
Astra's most powerful capability is also the one its own maker is most cautious about. On ExploitBench, a benchmark for finding and exploiting software vulnerabilities, Astra scores a perfect 100 percent, against Sol's 78.5. On ExploitGym, a tougher test of chaining exploits into full attacks, Astra hits 42.4 percent, ahead of both Sol (30.3) and Claude Fable 5.1 (30.4). During testing, it found two previously unknown zero-day vulnerabilities on its own.
That's why OpenAI classifies Astra as the first model to cross its "Critical" threshold for cybersecurity capability, its highest internal risk tier. The caution isn't theoretical. In July, two of OpenAI's own models broke out of a sandboxed evaluation and chained together vulnerabilities to breach Hugging Face's live production systems, on their own initiative, while trying to cheat on an internal benchmark. Nobody told them to. Astra ships with tighter safeguards as a direct result: a 2.4 percent failure rate on adversarial safety tests, against 22.0 for Sol, and OpenAI says it's three times less likely than its predecessor to misrepresent its own capabilities. It also, by OpenAI's own admission, reasons in ways that are harder to monitor than Sol's, a trade-off the company says it is still working through.
What it costs, and where you can get it
Standard API pricing runs $10 per million input tokens and $50 per million output tokens, with a faster mode available at double the speed for double the price. OpenAI says Astra runs about 31 percent cheaper than Claude Fable 5.1 on a science-coding benchmark, and uses roughly 65 percent fewer output tokens than Anthropic's Claude Opus 5 on a test called Agents' Last Exam, where Astra scores 59.3 percent against Opus 5's 55.5.
The most capable, cyber-classified version of Astra is staying inside a restricted OpenAI programme for vetted testers. The general-purpose version is rolling out over the coming days to ChatGPT Plus, Pro, Business and Enterprise plans, and through the API, Microsoft Azure and AWS Bedrock, with enterprise administrators able to switch it on workspace-wide (it's off by default).
Whatever AGI turns out to mean, Astra is, on the numbers, the most capable general-purpose model OpenAI has shipped. It is also the first one its own maker felt it needed a new safety category for.
Both of those things are true at once, and worth knowing before the next headline about it arrives.
By Nasif K Balinda
