On 12 June, OpenAI retired GPT-5.2 — Instant, Thinking and Pro — and moved everyone, mid-conversation, onto GPT-5.5. There was no long phase-out and not much notice; plenty of developers spent the day checking whether the prompts and integrations they had carefully tuned still behaved. For a small business this is the quiet risk nobody prices in: the model you built a workflow on is rented, not owned, and the landlord can change the locks. The fix is to build so the model underneath can be swapped without anyone in your business noticing. Here is the one rule I now design every client workflow around, and the ten-minute check that tells you whether a vendor switch would cost you a morning or a month.
I want to start with the unglamorous version of what happened, because the dramatic version is misleading. On 12 June, OpenAI turned off GPT-5.2 in ChatGPT and the API. Conversations that were running on it carried on, automatically, on GPT-5.5. If you only ever use ChatGPT through the app to draft the odd email, you may not have noticed a thing. That's the point: for casual use, a forced upgrade is mostly invisible and usually an improvement.
The people who felt it were the ones who had wired a specific model into something that runs every day. A prompt that was tuned to GPT-5.2's particular habits. An automation that expected the output in a certain shape. The categorisation step that was right 96% of the time on the old model now needs re-checking on the new one. None of that is broken, exactly. It just can't be assumed any more, and “can't be assumed” is expensive when a process touches your invoices or your customers.
The thing the launch coverage skips
Every model release gets covered as a sport: who's top of the table this week, what the benchmark chart says. The retirement of the previous model barely gets a paragraph. But for anyone running a business on this stuff, the retirement is the more important event. It tells you the real terms of the deal you signed up to.
Those terms are simple and worth saying plainly. You do not own the model. You rent access to it. The vendor can improve it, reprice it, or switch it off on their schedule, not yours. And increasingly that schedule is measured in months, not years. This isn't a complaint about OpenAI specifically; everyone in the market is moving at the same pace. June alone has brought a new Anthropic flagship and a forced migration to GPT-5.5. It's the tempo now, and it isn't slowing.
So the question for an owner isn't “which model is best.” It's “what happens to my business the morning my chosen model gets retired?” If the honest answer is “I'm not sure,” that's the thing to fix, and it's fixable in an afternoon.
The rule: own the layer above the model
Here's the principle I now design every client workflow around. The model is a component, not the product. The product is the workflow around it: the instructions, the quality checks, the human reviewing anything that touches a customer or a ledger. If all of that lives in your business and the model is just a swappable part underneath, a retirement is a non-event. If your “workflow” is a clever prompt that exists only in one person's ChatGPT history, you've built your house on rented land and you'll feel every move the landlord makes.
In practice, owning the layer above the model means three boring things are true:
- Your prompts and rules live somewhere you control — a document, a repository, a tool you own — not in chat histories scattered across staff accounts. If the model changes, you have one place to adjust and re-test, not a treasure hunt.
- You can name what “good” looks like. For any AI step that matters, there's a simple measure of whether the output is right — a sample you check against, a number you watch. Without it, you can't tell whether a new model made things better or worse; you're just hoping.
- A human owns the steps that touch money or customers. The AI drafts, suggests, sorts. A person approves the things that would be costly to get wrong. That single boundary turns a silent model change from a risk into, at worst, a slightly different first draft.
A property tech client of mine runs a daily job that reads enquiries and routes them. When GPT-5.2 went away, the change for them was nothing. The instructions live in their own system, there's a daily count of mis-routes they already watch, and a person eyeballs anything the model flags as unsure. The model swapped underneath and the business carried on. That's not luck. It's the layer.
What I would not do
The wrong lesson here is to go hunting for stability by freezing on an old model or refusing to touch AI until the market “settles down.” It won't settle down, and the old model gets retired anyway. You'd just be choosing to be surprised by it later. Pinning yourself to a specific version through the API buys you a little notice, not permanence; the deprecation notices still come.
The other wrong lesson is to over-engineer. A small business does not need a complex system that automatically swaps between five AI vendors. That's a solution to a problem you don't have, and it adds fragility of its own. What you need is to know where your instructions live, what good looks like, and who's holding the pen on the risky steps. Three sheets of paper, not a platform.
The reframe that makes this easy
Stop thinking of yourself as a customer of a model and start thinking of yourself as the owner of a process that happens to use one. Models will keep getting better, cheaper, and occasionally switched off without much ceremony. If your business is built so that's someone else's problem to manage and yours only to benefit from, every retirement is a free upgrade. If it's built the other way round, every retirement is a fire drill.
GPT-5.2 going dark on a Friday is a small story. The habit it should prompt is not. Spend ten minutes this week finding out which of your AI tasks would survive the next retirement untouched, and which one wouldn't. Fixing the one that wouldn't is the cheapest insurance in your whole stack.