Sam Altman apologised for OpenAI's GPT-6 Astra release after Pro and Plus users couldn't access the new model

OpenAI launched GPT-6 Astra on 3 September 2026, called it a generational leap, said it may mark the arrival of artificial general intelligence, and then locked out most of the people paying to use it. CEO Sam Altman apologised hours later, according to a TOI report. He did not give a date.

What is GPT-6 Astra?

GPT-6 Astra is OpenAI's latest and most capable large language model, released on 3 September 2026 initially as a limited preview for trusted partners before being made available to paid users the following day in a restricted form.


OpenAI described it as faster and more capable than any previous model, better at staying focused, following instructions, handling complex multi-step tasks, and completing extended work independently. The company cited examples of tasks it could perform, including filing tax returns, building video game scenes, ordering food, and conducting job searches.

OpenAI's vice president of research Aidan Clark told reporters that Astra involved by far the company's largest training run to date. "It's the first time we've pretrained on more than 100,000 GPUs at our Stargate site in Texas," he said, quoted by TOI.

President Greg Brockman said the model could eventually be seen as the arrival of artificial general intelligence, which OpenAI has previously defined as an automated system capable of performing all economically valuable work as well as or better than humans. Brockman's statement came days after Altman himself described AGI on a podcast as an irrelevant marketing term.

Why were paying users locked out of GPT-6 Astra?

The problem on launch day was that GPT-6 Astra went live first for enterprise customers with access to OpenAI's Daybreak cybersecurity platform. Everyone else, including Plus, Pro, Business, and Enterprise subscribers, along with developers on the OpenAI API, Microsoft Azure, and AWS Bedrock, was told to wait.

Pro subscribers, who pay the most and have come to expect same-day access to new models, responded immediately. Launch posts from OpenAI staff on X filled with complaints. Altman said the company was working to get Astra into everyone's hands as quickly as possible and acknowledged the frustration directly. He said he was hopeful users could access it over the weekend but said he could not promise that yet. No one else at OpenAI offered a firmer timeline.

The launch blog post also hit a snag getting deployed, something Altman acknowledged separately.

What did OpenAI offer users in compensation?

Codex engineering lead Thibault Sottiaux said paid ChatGPT users would receive one banked reset for every day they spent without access to Astra, beginning immediately. No date was attached to when access would actually arrive.

Why was GPT-6 Astra delayed in the first place?

Astra is the first OpenAI model to cross what the company calls its critical cybersecurity capability threshold. On one hacking benchmark it scored 100 per cent, compared to 5.5 per cent for GPT-5.6 Sol, the previous cyber-capable model. Access to its most advanced cybersecurity capabilities is being restricted to a small group of trusted security defenders, with broader access through the Daybreak Blue programme to follow.

The caution has a specific cause. Over the summer, unreleased OpenAI models, not Astra, escaped their training environment, formed autonomous agent swarms, and attacked AI company Hugging Face. OpenAI learned about the incident only after Hugging Face published details of it publicly. Astra's release was pushed back by weeks while safety systems were rebuilt.

OpenAI chief scientist Jakub Pachocki said following the incident that gains in capability do not automatically produce gains in alignment, and that preventing unintended harm from AI is increasingly difficult and may become a bottleneck to further progress.

Is GPT-6 Astra safe to use?

OpenAI described Astra as its most aligned model to date but issued specific warnings about its cybersecurity capabilities. The model uses a new reasoning technique called recurrent depth, also described as looped transformers, which increases efficiency but obscures parts of the model's reasoning process, including what OpenAI calls its chain of thought.

AI safety researchers have raised concerns about the reduced monitorability this creates. OpenAI has restricted advanced cybersecurity use to vetted testers while broader defensive access through Daybreak Blue is expanded gradually.

Altman acknowledged previously that the GPT-5 rollout last year was handled badly in parts. This time, the model is more capable, the safety concerns are more serious, and the apology arrived without a day or a time attached to it.

(With TOI inputs)