End of support for an AI model: what needs requalifying?

AI vendors · model lifecyclePublished October 15, 2026Méliantis team

In brief

AI vendors regularly retire their older models, with notice periods from 60 days to 6 months depending on the vendor. Each retirement forces you to retest the applications that depend on it, and the first difficulty is often knowing which ones.

Frequent, announced retirements

An AI model accessed through an API is not installed software: its vendor decides the date on which it stops responding. The major vendors publish their policy and a schedule.

Provider	Announced notice	Dated example
OpenAI	At least 6 months for generally available models, 3 months for their specialized variants, sometimes 2 weeks for "preview" versions	gpt-4.5-preview: announced April 14, 2025, shut down July 14, 2025
Anthropic	At least 60 days for public models	Claude 3.5 Sonnet: retired October 28, 2025. Claude Sonnet 4.5: retirement planned November 30, 2026
Google (Gemini on Google Cloud)	Retirement date published for each version	gemini-2.0-flash: retired June 1, 2026. gemini-2.5-pro and gemini-2.5-flash: retirement planned October 20, 2026
Microsoft (Azure OpenAI)	Retirement 18 months after launch, with at least 60 days' notice. Automatic upgrade on standard deployments	gpt-4o version 2024-05-13: retirement planned December 9, 2026

The Azure case deserves attention: on standard deployments, Microsoft automatically replaces the retired version, region by region. The change can therefore arrive without any action from your teams, and without a new test.

What needs requalifying

A new model answers differently, even with the same instruction. Before the retirement date, each application concerned must be checked on several points:

answer quality, on a representative set of test cases and with the same acceptance thresholds;
output format, when another application reads the outputs automatically;
guardrails : filters, refusals, behavior on sensitive requests;
cost and response times, which change with the model;
contractual terms and processing location, to be confirmed for the new model;
documentation : model inventory, validation files, user information.

What this touches in a project portfolio

Applications in productionAssistants, document classification, extraction: each was validated on a specific version.

Shared platformsAn internal AI gateway serving several teams passes the change on to every project that uses it.

Projects under constructionA project due to deliver after the retirement date was designed and tested on a model that will no longer exist.

Shared teamsThe same team often has to requalify several applications at once, to the same deadline.

Questions to ask yourself

Which models, in which version, does each of our projects use, including through an internal platform?
Which retirement dates fall in the next six months?
Do we have, for each application, a reusable set of test cases?
Which deployments are upgraded automatically by the vendor?
Which team requalifies what, and before which date?

See what this hits in your organization

Prévoq links this kind of change to the projects, teams and contracts of your organization, with the level of evidence of each link.

See the demo Request a discovery session

Sources

OpenAI, model retirements page · developers.openai.com · accessed October 15, 2026
Anthropic, model retirements · platform.claude.com · accessed October 15, 2026
Google Cloud, Gemini model versions and lifecycle · docs.cloud.google.com · accessed October 15, 2026
Microsoft, Azure OpenAI model lifecycle and retirements · learn.microsoft.com · updated July 24, 2026

This analysis presents the state of the texts and public sources at the date indicated. It does not constitute legal advice.