When the AI you rely on changes

AI services change on the vendor’s schedule, not yours. Several major providers publish deprecation and retirement schedules for their models. Whether a given change matters depends on what your organization relies on that model for.

Change is routine, not exceptional

Providers of AI models regularly release new versions and retire old ones. Several of them publish this on dedicated pages: lists of deprecated models, planned retirement dates and suggested replacements. The sources below link to some of these pages.

Retirement is only one kind of change. Terms of service, data-handling commitments, pricing plans, regional availability and default settings can change too, and so can the way an application you already use integrates AI.

Kinds of change to watch for

Mapped to the four distinctions in AI use:

  • AI system or capability: a model version is deprecated, retired or replaced; behaviour changes between versions.
  • Deployment Context: a licence, plan or agreement changes — including what happens to the data you send.
  • Application: a tool you already use adds, removes or changes an AI feature.
  • Reliance: your own use drifts — the AI is used for something more consequential than when it was first accepted.

Does this change matter to us?

The same change can be irrelevant to one team and important to another. A useful way to decide is to go back to why the reliance was accepted in the first place and ask whether that reason still holds.

  • Which of our uses depend on the thing that changed?
  • For each, what did we rely on — accuracy, availability, data handling, a person reviewing the output?
  • Does the change affect that reason directly, possibly, or not at all?
  • How critical is the decision or task that depends on it?

A proportionate response

Not every change needs a review meeting. Recording that a change was seen and judged irrelevant is often enough. A change that touches the reason behind a critical reliance deserves a closer look before continuing. The point is that the response should fit the change and the stakes, and that the judgement is written down.

Limitations

  • Vendor lifecycle pages describe each vendor’s own policies, which differ and change. Always read the current page rather than relying on this summary.
  • Not every vendor publishes a retirement schedule, and changes inside third-party applications are often announced less formally.
  • This article is general information, not legal or procurement advice.

Sources

  1. Deprecations — OpenAI API documentation. Last reviewed: .
  2. Model deprecations — Claude documentation (Anthropic). Last reviewed: .
  3. Model versions and lifecycle — Google Cloud documentation. Last reviewed: .
  4. Model deprecations and retirements — Microsoft Learn (Azure OpenAI in Foundry). Last reviewed: .