What Is Model Deprecation?

An AI model feels permanent until suddenly it isn’t. Model deprecation – the provider retiring a version and forcing a migration – is one of the most underrated risks in AI contracts. For most, it’s a footnote. For a regulated organization, it can become a crisis if the notice period isn’t written into the contract.

The definition

Model deprecation means an AI provider stops offering a particular model version and requires whoever uses it to switch to a newer one. It’s a normal part of the model lifecycle: new versions are released, old ones aren’t maintained forever.

The consequence for you depends on how you use the model. Sometimes a switch is an uncomplicated update. Sometimes – especially when the model’s behavior is baked into a business-critical or regulated process – it means a major redo. And what decides whether you make it in time is the notice period.

Why it matters even without formal requirements

Model deprecation is often seen as a problem only for strictly regulated industries. But even an ordinary business is affected, since a new model version doesn’t behave exactly like the old one. Prompts fine-tuned for one version can produce different answers on the next – sometimes better, sometimes worse, but rarely identical. If you’ve built a service where tone, format, or precision matters, it needs to be retested at every model change.

A concrete example: a customer service assistant has been tuned so its answers hold the right tone and length. When the underlying model is deprecated and replaced, the same instructions can suddenly produce longer or more formal answers. Without notice and a test period, you notice it only once your customers do. That’s why the notice period is a question for more than just those with a formal validation process.

The notice period varies dramatically

Here’s the core of the problem: how much notice you get differs enormously between providers and model types.

So the range goes from over a year down to a couple of weeks. Note in particular that some providers have no contracted minimum at all – then you’re at the mercy of a policy that can change. And observed patterns, like Copilot’s month or so, aren’t the same thing as a contracted commitment.

The consequence for regulated organizations

For organizations that must formally revalidate at every model change, the notice period becomes decisive. Revalidation means showing that the new model version meets the same requirements as the old one – work that takes time.

If the notice period is shorter than the revalidation window, a gap opens up: the model is deprecated before the new one has been approved. A concrete scenario: an organization with formal model validation gets two weeks’ notice that their model version is being retired, but their revalidation process takes longer than that. The result is that they’re forced to choose between staying on a discontinued model or running one that isn’t fully validated.

The solution is to run two model versions in parallel during the revalidation window – the old one keeps production running while the new one is validated. That requires you to have planned for it, and for the contract to provide enough notice.

Make the notice period a contract term

The conclusion is simple: treat the notice period as a negotiating point, not a footnote. Write a minimum notice period into the contract, plan for parallel operation where revalidation is required, and build the solution so a model change doesn’t force a total redo.

And since notice periods and deprecation policies change over time: check the provider’s current policy instead of relying on a number you read somewhere. What applied at your last contract may have changed.

Want to see how model deprecation connects to provider choice, residency, and operational reliability across a full AI project? There’s more on our AI page. Need help building an AI solution that can withstand a model change? Get in touch.