AI models disappear: build a retirement calendar before workflows break

Model retirements are easy to ignore until something stops working. An image campaign no longer renders. An internal training video gets stuck. A customer-support workflow that used to answer properly starts returning errors because someone still calls an old model ID.

This is no longer just an AI lab problem. When AI sits inside course material, marketing production, quote work, support, analysis or coding-agent workflows, someone has to own a boring thing: a calendar of which models are allowed, when they change, and who tests the replacement.

What happened: Claude and Gemini made model changes harder to ignore

Anthropic says in the Claude Platform release notes that claude-sonnet-4-20250514 and claude-opus-4-20250514 are now retired. Requests to those model IDs return errors. The recommended move is Claude Sonnet 4.6 and Claude Opus 4.8.

Source: Claude Platform release notes

Google made the same kind of signal clear in the Gemini API release notes on June 15. The Imagen IDs imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001 shut down on August 17, 2026. The Veo IDs veo-2.0-generate-001, veo-3.0-generate-001 and veo-3.0-fast-generate-001 shut down much sooner, on June 30, 2026. Google points video users to veo-3.1-generate-preview, veo-3.1-fast-generate-preview, or Veo 3.1 models through the Gemini Enterprise Agent Platform.

Source: Gemini API release notes

Those are dates and model names. Fair enough. But for an organization, this is really an operational planning problem. Do you have a workflow that says "use this model" somewhere in code, a no-code tool, a spreadsheet, a prompt template or an external partner's setup? Then the change needs to be planned like any other small system change.

A model calendar is simpler than it sounds

A model calendar does not need to be a heavy governance program. Start with a list. Each row should answer seven questions:

• Which model ID is used?
• Where is it used: app, script, agent, campaign, course, plugin, or supplier setup?
• Who owns the workflow?
• Is there a shutdown date or deprecation notice?
• Which replacement model should be tested?
• Which examples must be compared before the switch?
• Who approves production and rollback?

That gets you surprisingly far. The point is to find the dependency before users, students, customers or colleagues are waiting for output.

Image and video need more than a new name

For text models, a model change may show up as tone, format or cost. For image and video, the difference can be more visible: style, faces, text in image, length, format, timing, audio and brand feel.

That is why a marketing or education workflow should not simply replace veo-3.0-generate-001 with the next name and hope for the best. Run a few fixed test cases first:

• a typical campaign image
• a difficult prompt with brand requirements
• a video example with motion and timing
• an example where text in the image often fails
• a cost and waiting-time test

Save the old and new results side by side. The decision gets much easier when someone can see the difference, not just read a release note.

Put ownership close to the workflow

On the same day Google listed the media deprecations, Gemini Enterprise also added observability settings for individual Agent Designer agents in Public Preview. Metrics can go to Metrics Explorer and traces to Trace Explorer. That is a different product area, but the signal fits: AI workflows are getting owners, logs and operational trails, not just prompts.

Source: Gemini Enterprise release notes

If nobody owns the model, nobody owns the failure either. Keep the model calendar close to the real work. Marketing owns campaign images. Education owns course video. Support owns customer answers. An external partner can handle the technical switch, but the business still has to say which outputs are good enough.

A practical first week

Do not inventory every AI workflow on day one. Start with the five that would hurt most if they stopped on Monday morning. For each workflow:

1. Find the model ID or supplier setting.
2. Write down the owner and replacement candidate.
3. Run three old-vs-new test examples if possible.
4. Decide what pauses if the new model gets worse.
5. Put the next check in a calendar, not in someone's memory.

If this feels like a small, boring routine, that is a good sign. AI used in real work needs a few boring routines. Hammer can help build a lightweight Tool Forge process for model registers, test cases, fallback and ownership without turning it into a heavy IT project.