Your AI Feature Has a Retirement Date: A Migration Checklist for OpenAI and Claude Models

AI providers retire models on published dates. After a shutdown date, requests to a retired model fail, so a chatbot, document summary, or automation built in 2024 or 2025 can stop working without any change on your side. The fix is usually small, but it needs testing, and it is much easier to do before the date than after users report errors.

Dates to check (as published on 6 November 2026)

These dates come from the OpenAI API deprecations page and Anthropic’s model deprecations page. Providers add new entries regularly, so confirm the current list for the models you use.

• •23 October 2026 (already passed): OpenAI shut down gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o3-mini, o4-mini, gpt-image-1 and some early gpt-4o and gpt-4.1-nano snapshots. Features still calling these are already failing.
• •30 November 2026: OpenAI shuts down Agent Builder, the Evals platform, and the v1/prompts API with stored prompt objects. Anthropic retires Claude Sonnet 4.5 (claude-sonnet-4-5-20250929).
• •1 December 2026: OpenAI retires gpt-image-1-mini and gpt-image-1.5.
• •11 December 2026: OpenAI retires gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro.
• •6 January 2027: OpenAI retires the tts-1, tts-1-hd and gpt-4o-mini-tts text-to-speech models.
• •14 January 2027: OpenAI retires the gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini snapshots that many 2024–25 chatbots and automations use.
• •20 January and 26 February 2027: OpenAI retires several realtime, audio and transcription models, including whisper-1.

Anthropic also removed the temperature, top_p and top_k parameters in API version 2026-10-21. Requests that send them on that version return an error, so check request settings when you update a Claude integration, not only the model name.

1. Find every model your software calls

Search the codebase, environment settings, and any no-code automation tools for model names and API endpoints. Model names are often stored in configuration rather than code, and older integrations sometimes call a model through a plugin or a third-party service. Record each model, where it is used, and its retirement date.

• •Application code and environment variables.
• •Automation platforms, scheduled jobs, and webhooks.
• •Plugins, chatbot widgets, and vendor tools that call a model for you.
• •Stored prompts, agent configurations, and evaluation set-ups hosted with the provider.

2. Collect test cases before changing anything

Save a set of real inputs and the outputs your team currently accepts, including awkward cases such as incomplete requests or long documents. Remove personal or confidential data first. These examples let you compare the replacement model against today’s behavior instead of judging it on a few quick tries.

3. Switch to the recommended replacement and compare

Each deprecation notice names a recommended replacement. Run your test cases against it and check the results your users depend on: answer quality, output format, response length, speed, and cost per request. A newer model may follow instructions differently, so prompts and output parsing sometimes need small adjustments.

4. Move stored prompts and agent flows into your own code

If your product relies on OpenAI’s stored prompts, Agent Builder, or Evals, the 30 November shutdown removes the feature itself, not just a model. Move prompts into version-controlled application code or configuration, and rebuild agent flows and evaluations in tools you control. This also makes the next migration easier.

5. Release with a way back and keep watching

Release the change to a small group first or keep the previous setting ready until the retirement date, and watch error rates and user feedback. Then put a recurring check in your maintenance plan: providers usually give a few months’ notice, and a short review each quarter is cheaper than an emergency fix.

See how Estella Soft handles ongoing software maintenance

When to ask for help

If nobody on your team owns the code that calls the AI provider, or the feature was built by a developer who has moved on, a short migration check is a sensible start: an inventory of every model and API in use with its retirement date, one feature migrated and tested against your examples, and a written plan for the rest. Estella Soft can take on that work for software we did not originally build.

Explore Estella Soft’s AI integration services