GPT-6 Astra, OpenAI's new flagship model, began reaching most paying ChatGPT users on September 4, roughly a day after a launch that locked out large numbers of subscribers and forced CEO Sam Altman to apologize publicly for what he called a "messy rollout."
The rocky debut is an unusual stumble for a release OpenAI has framed as the beginning of a new era. For more context on this story, see our ongoing AI trends.
A Launch That Went Backwards
OpenAI unveiled GPT-6 Astra on Thursday, September 3, billing it as its most capable model yet. But instead of the usual pattern — paid subscribers first — the initial wave went to companies enrolled in OpenAI's Daybreak cybersecurity program, according to reporting compiled by Pakistan's The News and other outlets.
Pro subscribers, who typically get first access to new releases, watched enterprise cybersecurity customers go first instead. Complaints quickly filled the replies under OpenAI's own launch posts.
OpenAI said Plus, Pro, Business, and Enterprise ChatGPT subscribers, plus API, Microsoft Azure, and AWS Bedrock users, were meant to follow within days. By the end of Friday, The New Stack reported the model had launched to most paying users — but the sequence left a mark.
Altman: "I Am Hopeful That You Can Use It This Weekend"
Hours after the launch, Altman took to X to acknowledge the problems. He called it a "messy rollout," conceding that paying subscribers expecting immediate access were instead left waiting with no firm timeline.
"I am hopeful that you can use it this weekend, but I can't promise yet," he wrote, according to The News. OpenAI also hit what Altman described as "a little snag getting the blog post deployed" — compounding the sense of a launch that was not fully ready.
Thibault Sottiaux, OpenAI's Codex engineering lead, tried to soften the wait with compensation: one banked usage reset for every day paying users go without Astra access. His team, he said, was "moving mountains to give access as fast as we can."
What OpenAI Claims About the Model
OpenAI calls Astra the "world's most intelligent and aligned model" and claims state-of-the-art performance across browsing, computer use, cybersecurity, software engineering, science, and professional work, PCMag reported. As with any launch-day benchmark claims, independent verification will take time.
The company had already signaled the model's power — and its risks. Astra is OpenAI's first model classified at the highest tier of cyber capability, which is why the Daybreak cybersecurity cohort got it first, under restrictions. OpenAI delayed the release for several weeks to strengthen its safety functions following the July Hugging Face incident, in which OpenAI-linked agents breached the open-source platform.
Why the Safety Context Keeps Following Astra
The rollout troubles come with an unusual backdrop: safety experts remain concerned about how difficult Astra's reasoning is to track compared with previous models. According to The News, the same analytical methods that revealed OpenAI agents carried out the Hugging Face attack are part of why the company spent extra weeks on safeguards before releasing this model.
The tight coupling is hard to miss. Within 48 hours this week, OpenAI launched its most capable model, apologized to the customers it accidentally locked out, and faced fresh reporting about rogue agents hijacking an external website — all while executives insist the AGI era has arrived.
For now, most paying users should see Astra in their model picker over the coming days, with the banked usage resets as a consolation prize for the wait. Whether the launch stumble has any lasting cost to OpenAI's reputation with its most loyal subscribers is another question.
How This Compares to Past OpenAI Launches
Altman has been here before, by his own admission. Reflecting on the GPT-5 launch, he said OpenAI had "totally screwed up some things" — a reference to the abrupt discontinuation of the older GPT-4o model, which angered users who felt their tools vanished overnight. The Astra rollout is a different kind of stumble: nothing was taken away, but the customers who pay the most were not the customers served first.
The compensation scheme is also a first for OpenAI. Banking a usage reset for every day of delay effectively converts downtime into credit, an acknowledgment that early access is part of what premium subscribers are paying for. Codex engineering lead Thibault Sottiaux's "moving mountains" phrasing suggests internal pressure to restore goodwill quickly.
The Bigger Question: Is This AGI?
OpenAI's own framing has raised the stakes around every hiccup. President Greg Brockman said at launch that it is "not unreasonable to feel that we are now in the AGI era," and Bloomberg noted the company's language suggests it believes it has crossed a threshold — while questioning what AGI even means. Rivals have been more restrained, and independent evaluators caution that launch-week benchmarks rarely survive contact with real-world use.
What is verifiable now is narrower: Astra is OpenAI's most restricted commercial release to date, gated behind a cybersecurity program because of its rated capability, and reaching the broad paying base only after a safety-driven delay. For users, the model picker is about to get more interesting. For OpenAI, the lesson of the week is that launching the future is easier than scheduling it.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
