OpenAI chief executive Sam Altman apologized early on September 4, 2026, for what he called a “messy” rollout of the company’s new frontier model, GPT-6 Astra, one day after a staged launch left paying ChatGPT subscribers without the access they had expected. In a post on X, Altman told users the company should be able to begin a broad rollout to API customers and ChatGPT subscribers “in the near future,” starting, as usual, with Pro subscribers.
“When we screw up, we try to make it right,” Altman wrote in the same post. The apology capped a launch day in which OpenAI restricted Astra to a limited set of organizations while telling subscribers on Plus, Pro, Business, and Enterprise plans that access would arrive over the coming days.
What OpenAI Announced
OpenAI introduced GPT-6 Astra on September 3, 2026, describing it as the world’s most intelligent and aligned model. According to the company, Astra is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work, and it saturates FrontierMath Tier 4 with a 98% score, ARC-AGI-3 with a 99.9% score, and ExploitBench with a 100% score. OpenAI said the model combines years of research across pre-training, reinforcement learning, and alignment.
The announcement stated that Astra was rolling out that day to a limited set of organizations and would become available over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock. Astra usage is included within existing subscription allowances, with credits purchasable for additional usage, and Pro, Business, and Enterprise plans will also receive access to a GPT-6 Astra Pro tier. Enterprise access is off by default at launch and must be enabled by workspace administrators. In the API, the model is priced at $10 per million input tokens and $50 per million output tokens at standard processing speed.
Altman announced the launch on X at 7:49 PM on September 3, 2026, writing that OpenAI believes Astra is the best model in the world for computer use, professional work, science, coding, and cybersecurity. He said the company took extra time to meet the safety and alignment standards required for this capability level.
Subscriber Complaints and the Company’s Response
The decision to begin with a limited organizational tier left paying subscribers waiting, and Altman’s launch replies filled with complaints from users asking when access would arrive. At 7:53 PM on September 3, 2026, Altman posted that the company was working to get Astra into everyone’s hands as quickly as it could, acknowledging the frustration and asking for patience.
Thibault Sottiaux, a Member of Technical Staff at OpenAI who leads Codex, the company’s software engineering agent, announced a compensation measure at 11:12 PM on September 3, 2026. OpenAI will give one banked reset for every day a user lacks access to Astra on a paid ChatGPT plan, starting that day, with the first reset landing roughly three hours after his post. He said the team was “moving mountains” to expand access as fast as possible.
Asked by one user to pin down what “near future” meant, Altman replied at 1:09 AM on September 4, 2026: “i am hopeful that you can use it this weekend! but can’t promise yet.”
The launch day carried other stumbles. Minutes after announcing the model, Altman posted that the company had hit “a little snag” deploying the announcement blog post itself.
Safety Work Behind the Staged Release
OpenAI had signaled the staggered approach in advance. In a September 1, 2026 safety update, the company said it had determined that Astra meets the Critical cybersecurity capability threshold under its Preparedness Framework, the first model it has designated at that level, meaning it can find previously unknown security flaws and develop ways to exploit them across well-protected systems without step-by-step human guidance.
The company said it had delayed parts of Astra’s development and release over several weeks while it strengthened and tested protections against cyber misuse and unauthorized model actions. That work included pausing certain frontier training runs for two weeks after a security incident involving AI agents and the Hugging Face platform, and restarting a large frontier reinforcement learning run on August 28, 2026, once new safety and security requirements were in place.
OpenAI said access to Astra’s most advanced cybersecurity capabilities would be more limited, with that work initially available to a small group of alpha testers and expanding afterward through Daybreak Blue, part of the company’s Daybreak cyber defense program. At launch, the released model refuses more advanced cybersecurity tasks such as creating proof-of-concept exploits, while defenders can use it for secure code review and patching. The company also said it is deploying misalignment monitoring in production for Astra-class models, using classifiers that check the model’s reasoning and actions for unauthorized behavior, and warned that extra safety checks can sometimes slow, pause, or stop legitimate work.
In the safety update, OpenAI said its evaluations found Astra far more likely than its predecessor, GPT-5.6 Sol, to respect explicit safety restrictions and remain within its authorized scope, while also finding the new model’s written reasoning harder to monitor.
What Comes Next
As of Altman’s latest posts, OpenAI had not committed to a firm date for expanding Astra access beyond the initial organizations. Altman said the broad rollout to API customers and ChatGPT subscribers should begin in the near future, starting with Pro subscribers, and that he was hopeful subscribers could use the model over the weekend of September 5–6, 2026, without promising it. The banked-reset compensation for paid plans began on September 3, 2026, and continues for each day a paying user remains without access.
Credit: Source link
