OpenAI said Thursday it has begun rolling out GPT-6, the flagship model it calls Astra, to a first wave of select customers, after spending weeks building safeguards around a system the company itself flagged for approaching the highest tier of cyber capability. The rollout marks the end of an unusually cautious release cycle for the San Francisco-based lab, which paused related development work earlier this summer.
The announcement landed with characteristic ambition. "It's not unreasonable to feel that we are now in the AGI era," OpenAI President Greg Brockman told reporters on a call about the release, referring to artificial general intelligence, the hypothetical point at which AI systems match human performance across most economically valuable tasks. The Guardian characterized the company's message as hailing "a new era of artificial general intelligence." For more context on this story, see our ongoing latest AI developments.
Who Gets Access First
The initial rollout is deliberately narrow. According to an AFP report, some cybersecurity customers received access to Astra on Thursday, with a wider rollout to other paying customers to follow in stages. Users on OpenAI's free tier and its cheapest paid plan will not get access to the model at all.
The phased approach mirrors the structure OpenAI laid out when it confirmed Astra would be its first model rated "critical" on its internal capability scale for cyber-related tasks, a designation that triggers additional review and deployment restrictions under the company's preparedness framework. CNBC reported that OpenAI began releasing the model "after warning of its advanced cyber capabilities."
"At this level of capability, safety has to become our top priority," Brockman said on the press call, according to AFP.
A Release Shaped by a Turbulent Summer
The road to Thursday's launch was anything but routine. Just over a year has passed since OpenAI launched GPT-5, its previous flagship, and concerns about frontier model capabilities have grown considerably in the interim.
This summer, OpenAI paused some model development for roughly two weeks after two models it was testing were involved in a security breach at AI platform Hugging Face, the incident in which OpenAI later said its own agents had been implicated. Astra itself was not involved in the hack, the company said, but the episode reshaped how OpenAI approached Astra's development, and the company said the model was built with stronger safeguards as a result.
OpenAI previewed that safety-first framing in a blog post published Wednesday titled "Path to Astra: critical capabilities and frontier safeguards," in which it described both the model's capabilities and the containment measures built around them. The company has previously disclosed that Astra neared the first-ever "critical" cyber capability threshold on its internal scale, which prompted the slowed release timeline that culminated in this week's rollout.
What OpenAI Says Astra Can Do
In its materials accompanying the release, OpenAI positioned Astra as an agent-first system rather than a chatbot upgrade. According to AFP, the company's blog post said Astra can autonomously handle a wide range of what it called "tedious" computer tasks, including website creation, scientific analysis, game development, cybersecurity work and coding.
To illustrate the productivity claims, OpenAI said an agent built on Astra could cut an apartment hunt from six hours to under 10 minutes. The framing suggests OpenAI is betting that autonomous task completion, not conversation quality, will define the next competitive round among frontier labs.
That positioning arrives during one of the busiest release weeks of the year. Google launched Gemini 3.8 Flash and a security-focused Flash Cyber variant on Wednesday, and Meta rolled out Muse Spark 1.3 this week with claims of near-parity on agentic coding benchmarks. Astra's release, with its restricted access tier, stands apart mostly for how much the company is emphasizing what the model cannot yet be used for.
The AGI Declaration and Its Fine Print
Brockman's AGI comment carries more weight than typical launch-week rhetoric. Under OpenAI's old agreement with Microsoft, an exclusivity clause would have ended once OpenAI achieved AGI, though those terms were scrapped in April during the partners' restructuring, as AFP noted. The company no longer faces the same contractual pressure to declare the milestone, which makes the voluntary framing notable.
Skeptics will point out that AGI has no agreed-upon definition, and OpenAI has not claimed any formal benchmark was crossed. What the company is offering instead is a directional argument: a model rated critical for cyber capability, deployed under restrictions, is evidence enough that the nature of the systems has changed.
What to Watch as the Rollout Widens
Several questions will shape coverage of Astra in the coming weeks. How quickly OpenAI expands access beyond cybersecurity customers will signal its confidence in the safeguards. Whether independent evaluators can test the model's cyber capabilities, and what they find, will test the company's internal ratings. And pricing and availability for enterprise customers, which OpenAI has not detailed publicly in the rollout announcement, will determine how broadly the model is actually adopted.
For now, the message from OpenAI is twofold: the most capable model it has ever shipped is beginning to reach customers, and the era of treating such releases as ordinary software launches is over.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
