Model pinning (model versions and deprecation)
Calling an exact, dated model version instead of an alias like “latest”, so behavior does not change under you, and planning for its retirement date.
Aliases move to newer models, and a new model can follow your prompt differently even if it is better overall. Pinning makes behavior reproducible and lets you upgrade on purpose, after running your evals. Providers retire old versions on published dates, so a pinned model is a dependency with an end-of-life, like any library.
Related terms
- Prompt regression: An answer that used to be right goes wrong after a change to the prompt or the model.
- Evals: A repeatable test set that scores an AI feature, so changes can be measured instead of guessed.
- Model routing and fallback: Sending each request to the model that fits it, and to another one when the first fails.
