OpenAI's GPT-6 Astra is now in the hands of every paying ChatGPT subscriber whose plan includes it. Getting there took a day longer than promised, one public apology, and a scramble for computing power that the company still hasn't fully explained.
Astra was unveiled on Thursday, September 3, as OpenAI's most capable model yet. The catch: hours after the announcement, almost nobody could use it. API customers were waiting. Most ChatGPT subscribers were waiting. Eventually CEO Sam Altman showed up on X with an unusually plain mea culpa — "First, sorry for the messy rollout" — and a promise that broad access would start soon, Pro subscribers first.
Then Friday happened, and the story flipped. Pro, Enterprise and Business Premium users got access during the day, along with the API. By 6:30 p.m. Eastern, Plus and Business subscribers were in too — quicker than OpenAI's own engineers had predicted. Thibault Sottiaux, who leads engineering on Codex, sounded almost surprised announcing it, admitting the company's systems turned out to be "more scalable than we anticipated."
So what actually broke?
OpenAI hasn't said, at least not directly. Sottiaux came closest: the rollout leaned on "many novel systems" running at scale for the first time, while the company hurried to bring more compute online. The original plan was cautious — a limited group of organizations first, then the ChatGPT tiers, the API, Microsoft Azure and AWS Bedrock over several days. The demand apparently didn't care about the plan.
You can see why the hardware bill is steep. Astra ships with a 1.05 million-token context window and can produce up to 128,000 tokens of output. API pricing sits at $10 per million input tokens and $50 per million output — top-shelf rates for a top-shelf model.
Not everyone's invited
Here's the part that stings if you're on the cheap plan: ChatGPT Go subscribers, at $8 a month, don't get Astra. They don't get GPT-5.6 either, per OpenAI's pricing docs. The budget tier and the full-price tiers now live in visibly different worlds, and the cheapest ticket to OpenAI's best model remains the $20 Plus plan.
For subscribers who sat through the delay, there's a consolation prize: "banked resets," one for each day a paying user went without Astra access, counted from September 3. They're not bonus credits — they just let you refill your usage after hitting a cap, something OpenAI has done before with ChatGPT Work and Codex. Developers stuck waiting on the API side? Nothing so far.
The bigger deal here
A one-day delay isn't a scandal. What makes it interesting is what it exposed. Astra arrived wearing enormous claims — including persistent-agent abilities that let it grind away at tasks over long stretches — and while access lagged, nobody independent could check any of them. The early API testers who did get in have already hit surprises, like safety-triggered interruptions that look exactly like timeouts but aren't. If you're building production software on this model, that difference is not academic.
The launch also confirmed something about where AI competition has landed. When the hard part of shipping a flagship model is racking enough hardware fast enough, the model is only half the product. The other half is the power bill.
What's next
Watch three things. Whether API limits stay tight once the dust settles. Whether OpenAI ever gives a real post-mortem on what went sideways. And whether Astra's benchmark numbers survive contact with independent testing, now that the people doing the testing can finally log in. OpenAI didn't answer questions about the rollout or the API timeline. The apology's been made; the burden of proof is now on the model.
Image: panumas nikhomkhai, via Pexels
