OpenAI paused new Pro subscriptions. Astra demand ate the capacity, so the top tier is closed to anyone who wasn’t already holding a seat, and what’s on offer instead is a wait. I’ve been chewing on that for a day now, because it’s the first time I can remember a frontier lab saying, plainly, that the money is on the table and the compute isn’t.

You cannot buy your way past this one. The door is just shut.

A tier is a queue position wearing a price tag

We talk about subscription tiers like they’re product SKUs, the way a laptop config is a SKU. Pick the one that fits, pay, own the thing. But a model tier isn’t a thing you own, it’s a reservation against someone else’s cluster, and the reservation holds only as long as that cluster has room for you.

Most of the time the illusion is fine. Capacity is abundant, nobody’s counting, and the tier feels like a purchase. Then a launch lands with more demand than anyone modeled, and the abstraction peels off, and you find out that what you were actually buying was priority in a line.

I wrote something similar back in June when OpenAI rationed the GPT-5.6 rollout. Same shape. Different month.

So what breaks, exactly

Your work. Not the model. The model is fine.

The part nobody puts in the pricing page

The asymmetry bugs me. When a provider gates access, what you lose isn’t compute in the abstract, it’s the specific half-finished task sitting in front of you: the tab with the twelve rows you were extracting, the reply you were three sentences into, the dashboard you had filtered just right before everything stopped responding. Those things live in your browser, in a session you authenticated weeks ago, and they’re indifferent to whose GPUs are warm.

This is why I keep arguing that the browser is the durable layer and the model is the disposable one. A browser agent that runs in your logged-in Chrome, like Dassi, inherits everything your session already knows: your Gmail, your CRM, the internal dashboard behind SSO that no API exposes. None of that changes when a vendor closes signups. The context is local. The cookies are local. The page is already rendered.

What changes is which model reads it. And if that’s a dropdown, the worst case is you pick a different one and keep going, maybe grumbling for an afternoon that the replacement writes damn stiff emails. If it’s a hardcoded vendor relationship baked into the product you’re using, the worst case is you stop working and refresh a status page.

Swap the model, not the workflow

Dassi takes either your existing ChatGPT login or your own API key. Claude, Gemini, DeepSeek, whatever you’ve got credentials for. If you’re already paying OpenAI, that subscription works directly without touching a key form.

That’s the whole BYOK argument, and it’s less about saving money than people assume. It’s about the blast radius of someone else’s bad Tuesday. When Anthropic has an incident, you move to Gemini for the afternoon. When OpenAI caps Pro, you don’t care, because you were never buying access through a middleman whose own capacity deal just got squeezed. The key is yours, the browser is yours, and the agent sitting between them treats the model the way a power drill treats a bit.

Some models are better at long DOM reasoning. Some are cheaper for bulk extraction. Being able to test that yourself, on your actual tasks, in your actual tabs, is worth more than any benchmark table — and you only get to do it if switching costs you ten seconds instead of a migration.

Capacity is the new rate limit

I don’t think this is an OpenAI failure. Astra is popular, clusters are finite, and refusing money beats quietly degrading everyone’s experience.

But it does settle an argument. For two years the BYOK pitch sounded like something privacy people said at conferences. Now it reads as ordinary operational hygiene, the same instinct that makes you avoid a single point of failure anywhere else in your stack.

The irritating version of this story is the one where it happens again in October with a different logo on it, and everyone acts surprised a second time. I’d rather be the guy with a dropdown.