AI Model Deprecation: Planning Migrations Before Your Vendor Sunsets a Model

Frequently Asked Questions

01.How much notice do AI vendors give before retiring a model?

It varies by vendor and by model tier, which is the part founders miss. Anthropic commits to at least 60 days notice before retiring a publicly released model. OpenAI commits to at least six months for generally available models, at least three months for specialised variants, and as little as two weeks for preview models. Microsoft Foundry sets a standard retirement date 18 months out at launch and sends a notice at least 60 days before a generally available model retires. Your real planning window is the shortest notice in your stack, not the longest.

02.What is the most expensive mistake founders make when a model is deprecated?

Writing the model name into application code and prompt logic in dozens of places, so a swap becomes a code change rather than a configuration change. I did this in 2024 and it turned a routine replacement into three weeks of work across four services. The fix is an adapter layer that keeps model identity, prompt templates and output parsing in one place, plus an evaluation set you can rerun against a candidate model in an afternoon.

03.Does an AI model retirement create a compliance obligation?

For regulated entities in the EU, yes. DORA requires financial entities to hold documented and tested exit strategies for information and communication technology services supporting critical or important functions, and a model provider is such a service. The EU AI Act adds post-market monitoring duties for high-risk systems, which assume you can show the system still performs after a component changes. Practically, a model swap under either instrument needs a recorded evaluation comparing old and new behaviour on the same test set.