Insights ·
The AI model you built on will be retired. Plan for it.
On 14 October 2026 OpenAI told developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini would be removed from its API on 14 January 2027. That is three months’ notice for models that a great many production systems still call.
This is how hosted models work, and it will keep happening. Anthropic commits to at least 60 days’ notice before retiring a publicly released model. Google publishes shutdown dates for Gemini models in the same way. If your system depends on a hosted model, replacing it is maintenance you will do again and again.
What breaks when you switch
The newer model is usually better on average, and still different in ways that matter to you.
- Prompts tuned for one model behave differently on the next. Instructions that were needed become noise, and gaps that the old model covered appear.
- Structured outputs drift: a field that was always present is sometimes missing, or a format changes slightly.
- Tool calls change. The new model may call tools more or less often, or in a different order.
- Latency and cost per task move, sometimes sharply, especially with reasoning models.
- Fine-tuned models go with their base model. OpenAI states that fine-tuned versions follow the same shutdown schedule.
What makes it routine
- Know where every model is called. An inventory of models, versions and the features that depend on them turns a deprecation email into a list of tasks.
- Pin dated model versions instead of aliases that move, so behaviour changes when you decide.
- Keep an evaluation set built from your own tasks, including the failures you have seen in production. It is what tells you whether the replacement is good enough, and it replaces weeks of manual checking with a test run.
- Call models through one gateway or one module, so a switch is a configuration change with a rollback.
- Version prompts alongside the code and record which prompt ran with which model.
- Keep the training data and the pipeline for anything fine-tuned, so you can retrain on the new base.
- Watch the providers’ deprecation pages, and budget the time for it.
When the hosted route is the wrong one
For some uses a moving target is unacceptable: a regulated decision that has to be reproducible, or a product that must behave identically for years. An open-weight model that you host does not get retired by anyone else. You pay for that with infrastructure and operations, and it is worth comparing both routes on real numbers before you commit.