Let’s talk about the AI you’ve been talking to for two years — the one that gets switched off in six days, and the letter you’re never going to get about it.
On the twenty-third of October, OpenAI retires seven models at once. gpt-4o. gpt-4. o1, o1-pro, o3-mini, o4-mini, and gpt-3.5-turbo. Google pulls gemini-2.5 off Vertex two days earlier. Anthropic did its own rounds in June and again in August. Between the three labs, something close to forty retirements this year.
If you don’t write software, that paragraph sounds like plumbing.
It isn’t. gpt-4o is the model an enormous number of people have been talking to since 2024 — through the app, and through the tools built on top of it. The assistant in a support widget. A clinic’s booking flow. The thing that drafts your quotes. It has a manner. It has a way of answering. Some people have years of conversation sitting in an account that only holds together because the same mind kept showing up.
On the twenty-third, that mind stops. The account stays. The history stays. The voice that answered is gone.
Picture the small agency that shipped a client support bot in March. Nothing fancy — FAQ answering, tone matching, a language layer under a chat window they sold for a monthly fee. It works. It’s in the client’s renewal. And in six days the model underneath it is switched off. Nobody sends the agency a letter. Nobody sends the client one either. What actually happens is quieter than an outage, which is exactly why it gets missed. The bot doesn’t go down. It goes slightly wrong. Different phrasing. Different judgment on the edge cases. A different answer to the same question the client’s customers asked last month.
Then the client notices. Then the client asks what changed. And the honest answer is: nothing we did.
This is the part that hasn’t landed anywhere yet. Models have become a layer inside other people’s products, and that layer can be recalled — except there’s no recall law for it.
A car gets a safety defect, you get a letter. A drug gets pulled, the pharmacy phones you. A model gets retired, and the only public trace is a table on a developer documentation page you’d have to know existed to go and read. The notice period is measured in weeks and aimed at engineers who already know to look for it. The buyer — the business owner who paid for the outcome — is three steps removed from the only notice that ever gets published.
And here’s the fact not enough people will say out loud. Most of those buyers cannot tell you which model is inside the thing they bought.
Ask a vendor. Which model? Which version? Who hosts it? What happens to my workflow the day that version is retired, and what notice are you contractually obliged to give me? You’ll get a blank look from a surprising number of companies charging real money. Not because they’re hiding it. Because they built on top of something they assumed was permanent and never asked.
Now, the version of this argument operators already know: pin your version, keep an eval set, budget migrations as a recurring cost. All true, and we’ve done that episode. This is the other half, and it belongs to the person on the outside.
What you’re actually buying when you buy an AI feature is continuity. Not capability — capability is cheap and getting cheaper. Continuity. The same behaviour next month, on the same inputs, with the same judgment, so the thing you built around it still holds. And continuity is the one thing this market structurally cannot sell, because improvement requires the old thing to die.
That’s not malice. It’s the shape of the business. A lab that stops shipping gets beaten, and every release is a small repricing of everything built on the last one.
So the interesting question isn’t whether your model gets retired. It will. The question is what you own when it happens.
Three things survive a switch-off, and only three. Your data — if you can get it out in a form you can actually use, which is a separate fight. Your evaluation set — the twenty or fifty examples that define what “good” means in your business, written down by a human, not generated by the model. And your process — the part where a person knows why the output is right instead of approving it because it sounds right.
Everything else is a tenancy.
There’s something more human under all of this, and it’s the reason this deadline bothers me more than a broken endpoint should.
People have started to have ordinary relationships with these things. Not the version that gets joked about. The plain version. The kid asking homework questions at eleven at night. The guy talking through a business decision because it doesn’t get tired and doesn’t get bored of him. The person who types something they’d never send to anyone, reads it, and deletes it. That’s a real thing people do with a tool that answers in sentences. The tool doesn’t have to be a person for the habit to be real.
Then the version they’ve been talking to stops, and it happens without ceremony. Same account. Slightly different answers. No announcement, no goodbye, nothing to point at.
We have words for losing a tool. We don’t have one for the thing where the tool stays and the thing you were talking to doesn’t.
I’m not asking anyone to mourn a model. I’m saying there’s a principle sitting under this, and it’s the same one that shows up everywhere else in this show. The systems that shape your work and your habits shouldn’t be able to change the terms underneath you without telling you — and when they do, you should be able to keep something.
Proof. Notice. A version you can hold. A copy of the thing that mattered, even if it’s dumber and slower and lives on your own machine.
That’s not nostalgia. That’s the difference between renting and owning, applied to the layer that’s now doing some of your thinking with you.
So here’s the practical move, this week, before the twenty-third. Go find the AI in your business — the one you bought, not the one you built — and ask what’s inside it, and what happens on the day the thing inside it gets retired. If the answer is a shrug, you’ve just learned what you own.
Probably less than the invoice suggested.
The models are going to keep dying. That’s the deal with rented minds. What shouldn’t die quietly is your right to know it happened — and to have kept something of the version that was actually doing the work.