OpenAI Launched GPT-6 Astra to Headlines, Not to Paying Users

On 3 September, OpenAI launched GPT-6 Astra, the frontier model it had spent months teasing. The coverage was immediate and everywhere. The access was not. If you pay for ChatGPT — Plus, Pro, or a Business seat — there is a good chance you opened the app that day, went looking for the model in the headlines, and found it was not there for you.

Early access went first to a narrow group of enterprise and security partners. A day later, OpenAI’s chief executive Sam Altman apologised on X for what he called a “messy” rollout, and promised a broader release “in the near future,” starting, as usual, with Pro. And if you happen to live in the European Union, the wait is not measured in days: Astra shipped with no European data-residency option at all, and no date for one.

To be fair before the criticism, because fairness is the job: launching a frontier model safely and at scale is genuinely hard, staged rollouts are a reasonable way to manage capacity and risk, and Altman apologised quickly rather than pretending nothing was wrong. The problem is not that OpenAI was careful. It is that “launch” has quietly stopped meaning “you can use it,” and for some users it now means the opposite.

The launch where paying customers were last in line

The sequence is worth laying out plainly, because it inverts the usual order. OpenAI announced Astra, generating a wave of coverage about its new best model. Then it made that model available — first to vetted enterprise and security partners through a limited-access programme, framed around cyber-security work. The people who pay a monthly subscription for ChatGPT, the audience the launch was aimed at, were largely not in the first wave. They read that the model had arrived, then discovered they could not open it.

Altman’s apology the next morning was the tell. You do not apologise for a rollout that went to plan. He acknowledged the launch was “messy” and said the company should be able to begin a broad rollout to API customers and subscribers soon, beginning with Pro accounts. There was still no firm date for everyone else. In the meantime, OpenAI credited paying subscribers a usage reset for each day they spent waiting — a small, sensible gesture that also quietly concedes the point: the wait was real, and the people paying for it were owed something.

A launch is now a headline, not an availability date

This is the pattern worth naming, because Astra is only the latest example of it. An AI “launch” has split into two events that used to be one: the announcement, which is instant and generates the coverage, and the availability, which trickles out afterwards on a schedule the announcement rarely mentions. You are told the company shipped its best model. What actually shipped, on the day, was a press moment and a waitlist.

It did not help that Astra arrived in a week when OpenAI’s own reliability was already in question. The launch landed on the same stretch that saw a broad outage take ChatGPT down alongside its rivals — the kind of wobble that has, as we have written, stopped being news and become a pattern. A model you cannot reach because it has not rolled out to you, and a model you cannot reach because the service is down, feel much the same from the outside: the thing you pay for, not working.

You pay more for the model you can’t use yet

Then there is the price. In the API, Astra costs roughly $10 per million input tokens and $50 per million output tokens — about two and a half times the rate of its predecessor, GPT-5.6 Sol. So the headline model is both harder to get and more expensive to run once you get it. That combination is what sent users straight to the calculator during launch week, doing the maths on whether the frontier was worth the premium for everyday work rather than set-piece demos. A price rise is a defensible commercial choice. Announcing it in the same breath as a model most customers cannot yet touch is what makes it grate.

Europe doesn’t get it at all

For everyone in the launch queue, the wait is temporary. For the European Union, it is a wall. Astra shipped in two deployment flavours — a Global zone and a US data zone — and no EU data zone. OpenAI has not published a date for one. For an individual that is an annoyance; for a European business bound by data-residency rules, it is disqualifying. You cannot prove, per request, that data stayed in region if the model does not exist in region. Reports also note that projects pinned to EU data residency cannot use Astra’s faster “priority” or “fast” service tiers, and are limited to standard processing — assuming they can route to it at all.

There is a genuine steel-man here, and it deserves stating. EU data-residency and privacy rules are strict, building compliant regional infrastructure takes real engineering time, and shipping a half-finished EU option that quietly leaked data across borders would be far worse than shipping none. Not launching in the EU can be the responsible choice. But “responsible” and “silent” are different things. A frontier model that treats Europe as an afterthought, with no timeline and a downgraded set of options for those who do get near it, is of a piece with the pattern where Europeans reliably get the worse end of the deal — the market that, not long ago, got the ads first while the best features arrived later, if at all.

Safe for the vetted, off-limits for the rest

The two-tier shape of the launch shows up in the model’s behaviour, too, not only its availability. The public version of Astra refuses advanced cyber-security tasks — writing proof-of-concept exploit code, for instance — while looser safeguards are extended to vetted organisations through the same limited-access programme that got early entry. Automated monitors can interrupt a job mid-run if it looks to be heading somewhere prohibited, and that interruption behaves differently from an ordinary error, so it is not always safe to retry it blindly.

Gating genuinely dangerous capability behind vetting is a defensible piece of safety design, and the misuse pressure on a capable model is real; this is not the part to sneer at. But stack it on top of the rollout and the shape becomes clear. The fullest version of the model — earliest access, fewer refusals — goes to a vetted few. The ordinary paying user gets a more cautious version, later, at a higher price, and in Europe not at all. That is a defensible set of individual decisions that adds up to a frontier launch which, for the person it was marketed to, is smaller than it looked on stage.

What a launch should mean

None of this means Astra is bad, or that OpenAI acted in bad faith. The model may well be excellent, and some of this is the ordinary friction of shipping hard technology to millions of people at once. The narrower, fairer point is the one worth keeping: the distance between “we launched our best model” and “you can use our best model” has grown wide enough to mislead, and for European users it is not a distance but a closed door.

Until launches and availability line up again, a few habits help:

• Don’t rebuild your workflow on launch day. A model in the headlines may be weeks from your account; wait until it is actually in your plan before you depend on it.
• Check your own access page, not the announcement. Availability rolls out unevenly; the only reliable source for whether you have a model is your own settings.
• If you’re in the EU, assume no date. Treat “no EU data zone at launch” as “not available for compliant work until told otherwise,” and plan around the model you can actually use in-region.
• Weigh the premium before you switch. At roughly two and a half times the previous price, the newest model has to earn the upgrade for your actual tasks, not just the benchmark charts.
• Keep the old model in reach. The version you already rely on is the one that is available today; don’t retire it on the strength of a launch you can’t yet use.

The promise of a launch is a simple one: here is the thing, you can have it now. OpenAI is a long way from being the only company to have stretched that promise, but Astra is a clean example of how far it now stretches — a model announced to the world, apologised for the next morning, priced up, gated to the vetted, and, for an entire continent, not there at all. A launch should be a door opening. This one, for a lot of people, was a photograph of a door.