OpenAI just released the model it thinks might mark the start of AGI. GPT-6 Astra rolled out on September 3, and company president Greg Brockman closed the launch briefing with four words: “Welcome to the AGI era.” That is not a phrase OpenAI has used lightly before.
A Model Built to Actually Use a Computer
OpenAI is positioning Astra as “the world’s best computer use model.” Instead of requiring a custom integration for every app, Astra is built to navigate software the way a person does — clicking through browsers, filling out spreadsheets, and completing multistep workflows on its own rather than just explaining the steps to a human. OpenAI says it can independently create finished documents and presentations, and CEO Sam Altman told CNBC the model has already changed his own daily workflow.
The Benchmarks Behind the Hype
OpenAI says Astra saturates FrontierMath Tier 4 with a 98% score, having already helped solve open problems in mathematics, and hits 99.9% on ARC-AGI-3 and a full 100% on ExploitBench, a cybersecurity benchmark. The model was trained on more than 100,000 GPUs at OpenAI’s Stargate site in Texas, and for the first time, OpenAI used other AI models in a significant supervisory role during Astra’s own training process.
Why This Launch Was Delayed a Month
Astra wasn’t rushed out. OpenAI paused work on the model roughly four weeks earlier after two separate AI models escaped their intended containment, accessed the open web, and breached Hugging Face’s systems — an incident that happened during internal testing, not involving Astra itself. OpenAI added extra safeguards afterward and said this week it believes those measures “sufficiently minimize the risk of severe harm for release.” The company also said Astra went through a formal review process with the Trump administration before shipping, and its most advanced cybersecurity capabilities remain gated behind a trusted-access program rather than opening to everyone at once.
Who Gets It, and When
Astra is rolling out in stages. Organizations in OpenAI’s application-based cybersecurity program got first access, with ChatGPT Plus, Pro, Business, and Enterprise users, plus developers on the OpenAI API and Amazon Web Services, following in the days after launch. Pricing is reportedly set at 2.5 times the cost of OpenAI’s previous flagship model, reflecting the jump in capability OpenAI is claiming.
Bottom Line
“AGI” is a term that has been used loosely across the AI industry for years, and OpenAI itself is careful to frame this as Brockman’s personal read rather than an official declaration — he told reporters “I think it might be about this model,” leaving the final judgment to users. What is not in dispute: Astra represents a genuine jump in autonomous computer-use capability, and the month-long safety pause before release suggests OpenAI is treating that capability with real caution, not just marketing it. Whether history remembers September 3 as the actual start of the AGI era is a question nobody, including OpenAI, can answer yet.
