What is Model deprecation?

Model deprecation is the process by which an AI provider phases out a model. First it announces the plan. Later it retires the model, which means requests that name it no longer work or are redirected. The stretch between the announcement and the retirement date is the window you have to migrate.

Why providers do it

Running old models costs money. Each one needs GPUs, engineering support and safety upkeep. Providers want to focus that capacity on newer models, so they publish a schedule and retire the old ones. Many providers also close a model to new customers first, while existing customers keep access until the final date.

What a typical timeline looks like

Schedules differ by provider, so read the provider's own policy. The common stages are:

1. An announcement naming the model, the replacement and the retirement date.
2. A period where new sign-ups may be blocked while existing users continue.
3. The retirement date, after which calls fail or return an error.

Why it matters to you

A new model with a similar name is not a drop-in swap. Replacement models can differ in several ways that change your app:

• Settings. A newer reasoning model may fix values such as temperature or top-p sampling, or ignore the ones you set.
• Thinking and cost. A model that always reasons first uses more tokens, which changes your bill and your response time. See reasoning effort.
• Output. Format, tone and length can shift, which can break parsers and prompts tuned to the old model.
• Limits. The context window and rate limits can differ.

How to prepare

Inventory every place your product calls a model by name. Write a small set of real test tasks and keep them as evals. Run the old and new models side by side, compare quality, speed and cost, and switch gradually with a way to roll back. Do this before the deadline, not after calls begin failing.

If you use an open-weight model, you can keep running a copy yourself after a provider stops hosting it. That is one reason some teams keep a self-hosted fallback.

Related terms

Deprecation is a lifecycle event for a frontier model or any hosted model. It is the reason to keep your own evals: they tell you whether a replacement is actually good enough for your job.