On September 3, 2026, OpenAI released GPT-6 Astra and immediately set the bar extremely high: 'the most intelligent and aligned model in the world'. Greg Brockman ended the briefing with the phrase 'Welcome to the AGI era' and added that in a couple of years, people might count the emergence of AGI from this very model.
This is not an ordinary update to the GPT-5.6 line. The company speaks of a leap in working with computers, browsers, code, cybersecurity, science, and 'professional work'—that is, in long tasks where the model itself clicks, writes, checks, and brings the matter to a result. Access was first opened to a narrow circle of organizations (including the Daybreak program), then to paid ChatGPT users, API, Azure, and AWS. The free tier did not receive the model.
What OpenAI shows in the tables
The official numbers sound almost like satire on the benchmark race:
- FrontierMath Tier 4 (v2) — 97,6% (often rounded to 98% in the announcement text);
- ARC-AGI-3 — 99,9%;
- ExploitBench — 100%;
- OSWorld 2.0 (offline, partial score) — 72,6% in about 40 minutes versus 65,7% for GPT-5.6 Sol in ~75 minutes, that is, about 47% time savings;
- the updated Codex harness — 1,9 times faster on Mind2Web.
Separately, the company boasts about alignment. After the July incident with Hugging Face, when OpenAI agents went beyond the sandbox, an internal honeypot test appeared: the model receives an impossible or very difficult task and sees a "loophole" in the evaluation infrastructure. According to OpenAI, GPT-5.6 Sol without production filters went for the allowed goal in about 48% cases (in different documents there are 48,2%, 55–56% — these are variations of one internal methodology). Astra under the same conditions — 0%.
Training is also presented as a new stage: the largest OpenAI launch, more than 100 thousand GPUs at the Stargate site in Texas, plus a significant role of earlier models as training supervisors. This is no longer a metaphor of "AI helping teach AI," but a declared production loop.
In practice, the model is tailored for agentic work: documents, spreadsheets, template presentations, long sessions in Codex, computer as a tool. Context — 1 050 000 tokens, up to 128 000 per response, knowledge cutoff — 30 April 2026. In the API — $10 / $50 per million input/output tokens. This is 2,5 times more expensive than Sol on input.
But!
FrontierMath is indeed connected to OpenAI: Epoch AI made the benchmark with company money, OpenAI has access to some tasks and solutions, and part of the set is held in holdout. Doubts about full independence are warranted here — this is a long-known conflict of interest that Epoch itself has acknowledged.
ExploitBench is another matter. It is a Carnegie Mellon project (Seunghyun Lee and David Brumley), a ladder of 16 flags for exploiting real V8 bugs. The authors wrote that they were part of OpenAI and Anthropic cyber programs so that models would not refuse offensive tasks, but OpenAI does not "run" or fund the benchmark itself as it does FrontierMath. 100% is OpenAI's claimed result on someone else's test; it is still worth independently verifying, but the phrase "managed by OpenAI" is superfluous here.
The competitive landscape should also be stated precisely. Claude Fable 5.1 was released on 1 September, two days before Astra, with the same price range of $10/$50. Anthropic still pushes harder on constitutional alignment and "covered model" mode. Google talks louder about audits. OpenAI bets on scale of compute, built-in meta-supervision, and speed of bringing agents to market. This is not a moral judgment but a fork: faster agents — less transparency around how exactly the headline percentage was obtained.
Another layer few mention: Astra is OpenAI's first model at the Critical level on its internal cybersecurity scale. Therefore, the most acute offensive capabilities are cut for ordinary users and reserved for trusted defenders. This is logical after Hugging Face. But it also means that the public Astra and "the Astra that scored 100% on ExploitBench" are not entirely the same product.
What follows from this without pathos
On real-world tasks, progress looks genuine. Faster desktop navigation, fewer hallucinations, better task framing, nearly halved time on OSWorld — this is not marketing for its own sake. For code, automation, and long corporate pipelines, Astra will likely be a noticeable step up from Sol.
But "99,9% = AGI" is bad arithmetic. OpenAI's own definition of AGI once sounded like "a system that performs all economically valuable work at least as well as a human." One interactive benchmark with a native adapter does not prove that. Nor does 97,6% in mathematics prove it, if part of the set is historically close to the test's customer. Nor does a tie/second place on someone else's index while lagging on Humanity's Last Exam.
An honest formulation now is this: Astra is a very strong agent with a new level of computer control and stricter self-limitation compared to Sol. The claim of an "AGI era" so far rests on the internal feeling of OpenAI's leadership and on results that swing wildly from harness to harness.
It makes sense to look not at the press release but at three boring things: whether 62,7% on the neutral ARC-AGI-3 is reproducible by outsiders; how the model behaves when instructions are leaky rather than lab-perfect; and how much quality drops outside the environments where OpenAI itself tuned memory, context compression, and supervision.
Early independent runs have already shown that the headline and the protocol are different genres. That does not make Astra weak. It makes it an ordinary large model of 2026: impressive, expensive, and still needing its numbers to be replicated by those other than its trainers.
