On September 3, 2026, OpenAI president Greg Brockman stood in front of reporters and declared the arrival of the “AGI era.” Less than a day later, Sam Altman was on X apologizing to paying ChatGPT subscribers who still could not access the model that supposedly proved it. If you read our earlier look at whether GPT-6 Astra really earns that AGI framing, this is the part of the story that happened next, and it says almost as much about OpenAI as the benchmark scores do.
What Went Wrong: The GPT-6 Astra Rollout Locked Out Paying Users
GPT-6 Astra went live on September 3, but not for the people paying the most to use it. OpenAI gave first access to enterprise customers connected through its Daybreak cybersecurity program, while ChatGPT Plus, Pro, Business, and Enterprise subscribers, along with developers using the API, Microsoft Azure, and AWS Bedrock, were told access would arrive over the following days. Pro subscribers, who pay $200 a month and are used to getting new models first, were especially vocal, flooding OpenAI staff’s launch announcements on X with complaints within hours.
The optics were hard to miss: a company that had just finished arguing its new model represented a historic leap for humanity could not get that same model in front of the customers who had already paid for it.
Sam Altman’s Apology, in His Own Words
Altman posted on X early on September 4, opening plainly: “sorry for the messy rollout.” He followed with a short explanation of OpenAI’s general philosophy on mistakes, saying that when the company messes up, it tries to make things right, and confirmed that a broader rollout to API customers and ChatGPT subscribers would begin soon, starting with Pro subscribers as usual. He said he was hopeful people could get access over the weekend, but was careful not to promise a firm date.
What OpenAI Did to Make It Right
OpenAI backed the apology with a concrete gesture: affected paid subscribers began receiving a banked usage reset for every day they went without Astra access, starting immediately from launch day. By the evening of September 4, roughly 24 hours after the initial announcement, OpenAI confirmed that GPT-6 Astra had rolled out to all paying ChatGPT users with plans that included access, including Plus and Business subscribers. One OpenAI engineer, Thibault Sottiaux, posted that the company’s systems turned out to be more scalable than the team had anticipated, allowing the fuller rollout to happen faster than the cautious weekend timeline Altman had floated.
The Real Reason for the Delay: Astra’s Cyber Capability
The staged rollout wasn’t purely a communications misstep. GPT-6 Astra is the first OpenAI model to cross the company’s “Critical” cybersecurity capability threshold, meaning it can find and exploit flaws in well-defended systems without human help. That’s the same category of concern that led Google to restrict its own Gemini 3.8 Flash Cyber, which we covered in our piece on why Google locked its most capable security AI behind a vetting program. OpenAI’s Daybreak-first sequencing follows a similar logic, giving vetted security customers early access before the general public.
That reasoning is defensible on its own. What OpenAI clearly did not plan for was how it would look to have enterprise customers testing a capability most subscribers didn’t even know existed, while the same subscribers were left refreshing a loading screen for a model they were told was ready.
Why This Matters Beyond One Bad Launch Day
This is not OpenAI’s first public apology over a model launch. Altman previously admitted the company badly mishandled parts of the GPT-5 rollout the year before, after replacing the widely liked GPT-4o model triggered significant user backlash. Two consecutive launches now followed by a public apology is a pattern worth naming plainly: OpenAI keeps underestimating how much scrutiny a launch gets when hundreds of millions of people are watching for it to go smoothly, especially right after declaring it a historic moment.
What This Means If You’re a Paying ChatGPT User
- Check your account for the banked reset credits if you were on a paid plan during the delay, since OpenAI applied these automatically rather than requiring a support request.
- Don’t plan a product launch or client demo around a same-day new model release. This is the second time in two major launches that access has lagged behind the announcement.
- Treat launch-day capability claims with a short waiting period before you act on them. The most reliable read on a new model usually comes a few days after release, once real access and independent testing catch up to the announcement.
For the original reporting on Altman’s apology, The New Stack’s coverage includes the full timeline and his posts in context.
GPT-6 Astra Rollout: Final Thoughts
The GPT-6 Astra rollout ended up being a fitting coda to the AGI era debate it kicked off. A model good enough to prompt talk of a historic milestone still could not reach its own paying customers on schedule. OpenAI fixed the access problem within about a day and backed its apology with real compensation, which is more than some companies manage. But the gap between the announcement and the actual experience is exactly the kind of thing worth remembering the next time a launch event promises something historic.