Let’s talk about what happened yesterday, the twenty-third of October, and why almost nobody noticed.
On that one day, OpenAI shut down GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, o1, o3-mini and o4-mini. Six model families. Not deprecated. Not “we recommend you migrate.” Turned off. Every request to those names now fails.
About six months of notice, if you were paying attention. And here’s the part that should make you sit up: most people weren’t.
There’s a study that came out in September, out of Myongji University in Seoul, that went and looked at what actually happens when a model retires. Twenty-two thousand commits across seventeen thousand repositories. Five thousand of them matched to an official vendor shutdown announcement.
The finding is brutal. Roughly eighty-two percent of migrations away from a retired model were committed after the shutdown date. After the app had already started throwing errors. After a customer had already seen the failure.
Eighty-two percent. Not a fringe. Not the sloppy end of the market. That’s the baseline behaviour of professional teams.
Picture the Tuesday morning.
Somebody in a small firm in Pretoria built a quoting tool eighteen months ago. It reads a client’s email, pulls out the scope, drafts a quote in the house format, and a human checks it. Saved them maybe six hours a week. It runs on a model name that was current when they built it.
Nobody on that team is thinking about deprecations, because nobody on that team is paid to think about deprecations. There’s no platform team. There’s a developer who also does the invoicing integration and also answers the phone when the printer breaks.
The email from the vendor arrived in March. It went to an address that a contractor set up. It sat there. Then one Tuesday in October, the tool starts returning errors on every third request, and everyone stands around saying the same thing: nothing changed on our side.
Nothing changed on their side. Everything changed on the other side, on a schedule that was published six months earlier, in a place nobody was watching.
Here’s the mechanism, and it’s worth understanding properly because it’s not an accident.
A commercial model isn’t a library you install. You can’t pin it and keep it. The moment the vendor flips the switch, the thing your software depends on stops existing. Not breaks in an interesting way. Stops.
And the notice you get depends entirely on which tier of customer you are. Generally available models: at least six months. Specialised variants: three months, minimum. Preview models: as little as two weeks. Anthropic’s public commitment sits at sixty days. Google publishes an earliest shutdown date and doesn’t always publish when they announced it.
Trackers now put model lifecycles at six to twelve months, down from eighteen to twenty-four. So the window between “this is the best model in the world” and “this model no longer exists” is now shorter than a typical enterprise procurement cycle.
Now here’s the interesting bit, the bit that turns this from an annoyance into a pattern.
Eighty-two percent of teams migrate late. But only eight percent of migrations switch provider.
Read that again. Nearly everybody is doing forced, unplanned, expensive rework — and almost nobody uses that moment to leave.
Which means deprecation isn’t a cost the vendor pays for shipping fast. It’s a cost the customer pays, on the vendor’s schedule, and it renews the relationship rather than threatening it. You don’t shop around at eleven at night with production down. You take the recommended replacement, you change the model string, you fix the prompts that broke, and you go back to bed grateful.
The study found model identifiers hard-coded in ninety-four percent of the applications that migrated. Ninety-four percent. The model name is welded into the code, which means the abstraction you promised yourself you’d build never got built, because it worked fine and there was no deadline.
Then look at what migration actually costs. For a prompt-only app, a median of six added lines. Sounds trivial. For a fine-tuned one, closer to seven hundred. That’s the difference between changing a name and rebuilding a capability. And notice the trap inside that: the deeper you go, the more it costs to move, and the deeper you go the more the vendor’s tooling has already made the jump feel natural.
There’s one more detail that gives the whole thing away.
The same model can have two different death dates depending on which door you came in through. Claude Sonnet 4 was retired on Anthropic’s own API in June 2026. On Amazon Bedrock the same model ran until the fourteenth of October. Opus 4.1 died on the Anthropic API in August; on Bedrock it lives until the eighth of January.
The kill date doesn’t belong to the model. It belongs to whoever owns the platform you’re calling. Which tells you exactly who owns the relationship.
So what’s the operator lesson, because this isn’t a story about OpenAI being sinister. Every vendor in this market does a version of it, and most of them are telling the truth when they say the new model is better. The problem isn’t intent. It’s that “we improved the product” and “your product broke” are the same event, and only one of them shows up in your calendar.
Four things, and they’re boring, which is why they work.
One. Get the kill dates out of the docs and into your own system. The models endpoint returns a shutdown date for every model that has one. Pull it, put it in a table, and give each row a human owner. A notice with no owner is not a notice — that’s the eighty-two percent, right there. It isn’t ignorance. It’s unassigned.
Two. Own an eval set before you need one. Not a benchmark. Fifty to two hundred cases from your actual work, with your actual edge cases, and the answers a human on your team would sign off. When the replacement model arrives, you run it in an afternoon and you get a straight answer about whether your thing still works. Without that, “we migrated” means “it compiled.”
Three. Keep a second provider on a live, tested path. Not a slide that says model-agnostic. A branch that actually runs, quarterly, with real traffic on it. The reason vendors can keep doing this is that eight percent of you leave. Be annoying about it.
Four. Price the rework into the deal. If a model has a twelve-month life, your true cost isn’t tokens. It’s tokens plus a migration every year, paid in your engineer’s evenings. Put that line in the budget and the “cheap” model stops looking cheap.
And then the thing underneath all of it.
We’ve spent twenty years being told software is an asset. You buy it, you own it, it sits on your balance sheet. What’s being sold now is closer to a lease with a demolition clause. The intelligence is rented, the interface is rented, and the schedule belongs to someone else’s product roadmap.
The asset you actually own is the thing you built around it: the workflow, the customer relationship, the judgement about what a good answer looks like, and the eval set that encodes it. That’s yours. That survives a vendor.
A small business should be able to lose a supplier without losing its memory. If it can’t, it never had one — it had a landlord with a friendly interface.
So go and look at what your product actually calls. Find the model names. Find the shutdown dates. Find out who would notice if one of them disappeared on a Tuesday.
If the answer is nobody until the customer does, you already know what next week’s job is.
The machine will keep shipping. That’s its job.
Knowing what it costs you when it does — that’s yours.