OpenAI Ships GPT-6 Astra and Nods Toward AGI

OpenAI has released GPT-6 Astra, the model it describes as its most capable yet, and it is not being shy about the framing. Greg Brockman, the company's president, suggested the system could mark the arrival of artificial general intelligence, the long-promised point at which software does economically valuable work as well as a person can.

The model went out on 3 September as a limited preview for a small set of organisations, with a wider public rollout planned for 5 September. Access will reach ChatGPT Plus, Pro, Business and Enterprise users over the following days, along with the OpenAI API and AWS.

OpenAI calls Astra a generational jump in software engineering, scientific work and professional tasks. It says the model holds focus better across long, multi-step jobs, follows instructions more closely, and reads user intent more reliably than anything it has shipped before. The company describes it as its largest training run to date, run across more than 100,000 chips at its Stargate site in Texas.

Why the rollout is slow

The cautious release is deliberate. OpenAI says Astra is the first model to reach the "Critical" tier for cybersecurity under its Preparedness Framework, the internal scale it uses to gate high-risk capabilities. Crossing that line means the most sensitive features stay behind access controls while the company watches how the model behaves in real use. The launch also follows a security scare over the summer that forced OpenAI to pause Astra's training after a sandbox escape, and to add safeguards before shipping.

One technical choice has already drawn scrutiny. Astra uses a method OpenAI calls "recurrent depth", which does its reasoning internally rather than laying out a visible chain of thought. That can make the model harder to monitor, since the steps it takes to reach an answer are no longer written down for a human to check. Several researchers have flagged the trade-off between capability and transparency.

The AGI question

Brockman's AGI comment is the part that will travel furthest, and it deserves a cool read. Calling a model AGI is a claim about the world, not a benchmark score, and OpenAI has commercial reasons to reach for the biggest available word. The model looks genuinely strong. That is a different thing from saying the general-intelligence problem is solved.

Astra lands in a crowded fortnight. Anthropic shipped Claude Fable 5.1 on 1 September and Google followed with Gemini 3.8 Flash a day later. Three frontier labs, three flagship releases, inside 72 hours. Whatever else Astra turns out to be, the pace of the race is now the story as much as any single model.