Remember when the hardest part of buying software was deciding whether you actually needed it? You’d weigh the monthly cost, poke at the free tier, maybe expense it and hope nobody asked. The vendor’s job was to convince you. Your job was to say yes or no.
That relationship just inverted. OpenAI has paused new sign-ups for its $200-a-month ChatGPT Pro plan because demand for its new Astra model has outrun what the system can serve. Existing subscribers keep their access. Everyone else gets to wait. Two hundred dollars a month, cash in hand, and the answer is not right now.
What actually happened
The pause is a capacity decision, not a pricing one. Thibault Sottiaux, who leads product for Codex and ChatGPT, described it as the least disruptive option available, saying the company wanted to take the smallest step that would let it keep serving the broadest set of users. Someone else at OpenAI put the demand curve more bluntly: they’d been through steep growth before, but nothing like this.
Worth remembering that this is a company that already went well beyond its original Microsoft Azure arrangement for compute. It added CoreWeave. It launched Stargate, a $500 billion four-year infrastructure build. That is an enormous amount of planned capacity, and the demand curve still bent faster than the buildout.
The part that matters if you build with these tools
I review toolkits for a living, which mostly means finding out where things break under real conditions rather than in a demo. This event is a useful data point about a failure mode most people don’t test for: your vendor being unable to sell you more of what you already depend on.
Most procurement thinking assumes supply is elastic. If the tool works, you buy more seats. If the team grows, you scale the plan. That assumption held for basically every SaaS product of the last fifteen years, because marginal cost was close to zero and the vendor wanted your money more than you wanted their product.
Frontier models don’t work that way. Every additional heavy user consumes real compute that has to physically exist somewhere. When the top-tier plan sells out, it means the constraint is silicon, power, and data center floor space, not sales capacity.
So a few things follow, and they’re all practical:
- Grandfathered access has real value now. If you’re already on Pro, you’re holding something that new money can’t buy today. That changes the math on whether you cancel during a slow month.
- Availability belongs in your evaluation criteria. Benchmark scores tell you what a model can do. They tell you nothing about whether you’ll be able to onboard five more engineers next quarter.
- Single-vendor dependency just got more expensive. Not in dollars, in optionality. If your workflow assumes one specific model at one specific tier, you’ve tied your roadmap to someone else’s supply chain.
- Abstraction layers earn their keep. The tools that let you swap providers without rewriting your prompts and plumbing looked like over-engineering a year ago. They look like sensible insurance now.
Give credit where it’s due
Pausing new sign-ups is the honest version of this problem. The dishonest versions are the ones I usually run into during testing: silent rate limits, quiet quality downgrades where you’re routed to a smaller model without being told, or response times that quietly triple while the marketing page still promises the same experience.
I’d rather be told no at the checkout page than discover six weeks in that the thing I bought performs like a cheaper thing. Protecting existing subscribers over booking new revenue is the right call, and it’s not the call every company would make.
What I’d actually do about it
If you’re running production work on ChatGPT Pro, treat your current access as an asset with a switching cost attached, and document what specifically you’d lose if it went away for a week. If you were planning to add seats, start that conversation earlier than you think you need to.
If you’re evaluating options right now, this is a decent moment to test the alternatives you’ve been meaning to test. Not because Astra isn’t worth the money — I have no reason to think it isn’t, given how hard people are trying to buy it — but because knowing your fallback works is cheap right now and expensive later.
The broader shift is simple enough. AI tooling has moved from a buyer’s market to something closer to a supply-constrained one, at least at the frontier. That doesn’t last forever. Capacity gets built and constraints ease. But it’s the condition you’re operating in today, and planning around it beats being surprised by it.
đź•’ Published: