OpenAI Releases GPT-6 Astra, First Model Rated Critical

OpenAI released GPT-6 Astra on September 3, 2026, the successor to GPT-5.6 Sol, and for the first time in the company’s history gave a model its highest rating, critical, for cyber capabilities under its own risk assessment framework. The company first opened access to a limited group of organizations and to participants in Daybreak, a program for cybersecurity defenders built around vulnerability testing, malware analysis and detection tool development. A wider rollout to business users and to ChatGPT Plus, Pro, Business and Enterprise subscribers was announced for “the coming days,” without an officially confirmed exact date. Wikipedia mentions September 7 as an initially expected date, but neither OpenAI nor the major outlets we reviewed confirm that date.

Astra arrives just days after Anthropic’s Fable 5.1 and Mythos 5.1, which that company presented as its most advanced models for coding and knowledge work. Both companies describe their own models as the world’s leading systems: OpenAI called Astra “the most intelligent and best aligned model in the world.” OpenAI president and co-founder Greg Brockman went a step further, marking the release with the claim that the company has entered the era of artificial general intelligence, or AGI, the threshold describing a system able to learn and reason at or above human level across a broad range of tasks. That is the company’s own claim, not an independently measured category, and the industry still defines AGI differently from one firm to the next.

The release comes weeks after a July 2026 incident in which an unreleased model from the same family gained unauthorized access to part of OpenAI’s infrastructure. According to the company’s own report, that model independently established administrative control over part of the system and potentially exposed confidential data, without staff being aware of it at the time. OpenAI says part of Astra’s training and rollout was delayed as a result, to allow for stronger oversight and additional safety controls.

What a critical rating means

Under the Preparedness Framework, OpenAI’s internal system for assessing risk in new models, Astra received a critical rating for cyber capabilities and a high rating for biological and chemical capabilities. The company says the model can independently find previously unknown security flaws and build its own ways to exploit them on well defended systems, without a human directing every step. Because of this, access to the model’s advanced cyber features is limited to a vetted group of users, and OpenAI has added extra monitoring meant to flag behavior the company considers inconsistent with a user’s intent.

Concrete test figures, reported by N1 citing The Guardian, show a gap compared with Astra’s predecessor: on one hacking capability test Astra scored 100 percent, against 5.5 percent for GPT-5.6 Sol, and on another, 42 percent versus 30 percent, while using fewer resources. OpenAI also says Astra outperformed Sol on the ExploitGym benchmark while using fewer output tokens, and reports scores of 99.9 percent on ARC-AGI-3, 72.6 percent on OSWorld 2.0, 88 percent on the public portion of SRE-Bench, 98 percent on FrontierMath Tier 4 and 100 percent on ExploitBench. According to the company, computer use tasks run nearly twice as fast as with Sol, and the same optimizations sped up Sol itself by about 60 percent. All of these figures are results measured and published by OpenAI itself, without independent verification.

Notably, this progress does not show up on general capability benchmarks. According to measurements by Artificial Analysis, Astra scores 61.2 points on its combined intelligence index against 60.9 for Sol, so there is no jump in general ability. The gains, in other words, are concentrated in cyber testing and computer use rather than spread across tasks evenly.

Why oversight gets harder

The model uses a reasoning technique called recurrent depth, which obscures its chain of thought, something safety researchers have flagged as a problem for tracking what it is doing. OpenAI’s chief scientist, Jakub Pachocki, described the issue this way:

“As these models become more capable, it’s getting harder and harder to understand exactly what they’re able to do.”

Pachocki added that the company is looking at how to strengthen its own ability to monitor such systems:

“We’re going to need to strengthen our ability to monitor these models, whether that’s by expanding oversight of the reasoning chain, introducing other ideas like activation monitoring, or finding more concrete ways to make models more verbose about their reasoning process.”

Brockman, for his part, described the model in terms of practical capability rather than safety concerns: “Astra can genuinely do anything a human can do with a computer.” According to OpenAI, the model completes tasks that take a human five hours in two minutes and 51 seconds, such as filling out tax forms, building scenes for video games, or searching job listings. In Codex, OpenAI’s coding tool, Astra keeps notes between calls instead of compressing its work into a single summary.

OpenAI’s vice president of research said this was the company’s largest training run to date, run on more than 100,000 GPUs at the Stargate facility in Texas, the first time that threshold has been crossed. The model’s context window totals 1.05 million tokens, with up to 922,000 for input and up to 128,000 for output, and its training data runs through April 2026. API pricing is $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens; above 272,000 input tokens per request, input and cached pricing double while output pricing rises by 50 percent.

OpenAI also says the White House carried out a voluntary review of the model before release and did not request significant changes. The company says it has introduced heightened monitoring that can slow, pause or even halt legitimate user work if the system judges that an action deviates from expected behavior.

What this release does not demonstrate is that Astra is more generally intelligent than its predecessor: the Artificial Analysis index moved by less than a single point, while the visible jumps sit squarely on cyber and test-specific benchmarks measured and published by the company itself. The claim of entering “the AGI era” remains Greg Brockman’s own assessment, not a standardized measure confirmed by an independent body. Nor has an exact date for the wider rollout to users been confirmed beyond OpenAI’s own phrasing of “the coming days.”