“Sorry for the messy rollout”: OpenAI launched GPT-6 Astra, but developers are locked out

OpenAI said GPT-6 Astra was live on Thursday. For many developers, though, that announcement landed like a closed door. The model’s API docs and pricing are already public, but the endpoint itself is still making its way out, and OpenAI has not said exactly what went wrong.

That leaves a strange gap: the company has published enough detail to tempt builders, yet not enough access for them to test the thing on their own workloads. Astra’s specs are not small ones. OpenAI says it has a 1.05 million-token context window, can produce up to 128,000 output tokens, and charges $10 per million input tokens and $50 per million output tokens. Benchmark results were presented as strong, but independent checks are on hold until access catches up.

Sam Altman tried to blunt the frustration with an apology on X, calling the launch a “messy rollout.” He said broader access should start “in the near future,” beginning with ChatGPT Pro subscribers. OpenAI’s engineering lead for Codex, Thibault Sottiaux, said the rollout would take “a few days” and that the company is bringing new systems and additional compute online while “many novel systems” run at scale for the first time.

OpenAI’s original plan was broader than what users are seeing now. The company said Astra would first go to a limited group of organizations, then to ChatGPT Plus, Pro, Business and Enterprise users, the API, Microsoft Azure and AWS Bedrock over the coming days. But the delay matters more because Astra’s API also introduces safety-triggered interruptions that can look like timeouts, which is exactly the sort of thing developers need to understand before they build production systems around it.

To soften the blow for paying ChatGPT users, OpenAI is handing out what it calls a banked reset for each day they go without Astra access, starting September 3. That’s not extra credit and not a permanent bump in limits. It’s just a reset you can save for later. Developers waiting on the API get no similar promise yet, which is a neat summary of where things stand: the marketing page is ready, the model is not.