Altman apologised for the Astra rollout hours after calling it the AGI era

OpenAI called GPT-6 Astra a generational leap in capability on Thursday and the start of the AGI era. Within hours Sam Altman was apologising for a messy Astra rollout, because the paying customers who expected the model on launch day did not get it.

Who got it first

Access went to a set of enterprise customers, specifically those already inside Daybreak, the company's cybersecurity programme. OpenAI told everyone else, from Plus and Pro to the API, Azure and AWS Bedrock, to expect it over the following days.

Pro subscribers pay the most and are used to being first. The comment threads under OpenAI's own launch posts on X filled with people asking where the model was.

“We are working towards getting Astra in everyone's hands as quickly as we can. I know it is frustrating and I appreciate the patience. It should be quick.”

— Sam Altman, CEO of OpenAI, The Verge, 4 September 2026

Quote source: The Verge, 4 September 2026

No date, a credit instead

Altman gave no date. He said the company was working towards getting Astra into everyone's hands as quickly as it could, that he knew it was frustrating, and that it should be quick. Pressed further, he wrote that he was hopeful people could use it over the weekend but could not promise.

Codex engineering lead Thibault Sottiaux offered compensation instead of a date: one banked reset for every day a paid ChatGPT subscriber goes without access, starting that day.

The ordering is the actual story

Read the queue against what OpenAI published the same morning and it stops looking like a capacity failure. Astra is the first model the company rates Critical for cyber under its Preparedness Framework, and it went first to the programme built around defending infrastructure.

Our piece on the safety overview has the numbers behind that rating: 100% on ExploitBench, and 91.5% of cyber jailbreak attempts refused. A model scoring like that is a strange thing to hand to the largest possible audience on day one. That is a defensible order of release. Nobody at OpenAI presented it that way, so it landed as a queue problem instead.

The same programme collected the money as well. OpenAI committed $1 billion to Daybreak this week, in credit against its own products, and the organisations in it were the ones holding Astra while everybody else waited.

The rest of the day matched. Altman said there was a snag deploying the blog post. Safety researchers objected after the company acknowledged that Astra's reasoning is harder to monitor than earlier models, which sits against its own commitment to withhold scaling if monitoring degrades.

This is the second flagship launch in a row to open with an apology. After GPT-5 last year Altman said the company had totally screwed up some things on the rollout. The pattern is long enough that the apology reads as part of the launch plan rather than a reaction to it.

None of this should be read as personalized investment advice.