DEV Community

Sam
Sam

Posted on

Your AI model is retiring. Here's what actually breaks.

Your AI model is retiring. Here's what actually breaks.

Two big deadlines are close. Anthropic retires claude-sonnet-4-5-20250929 on Nov 30, 2026, and requests to it will fail after that. OpenAI shuts down gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro on Dec 11, 2026.

This month I sent 20 small pull requests to open-source apps still using models like these, and 8 have been merged so far. The lesson from all of them: changing the model name is rarely the whole job. Here's what else breaks, and the fix for each.

1. Find every place the model is named

Search the whole repo, not just the API call: code, environment variables, config files, model dropdowns, tests, docs and examples. Search for both forms of the name, like gpt-5-mini and gpt-5-mini-2025-08-07, or claude-sonnet-4-5 and claude-sonnet-4-5-20250929.

2. Claude Sonnet 4.5 to Sonnet 5.5

Anthropic's replacement is claude-sonnet-5-5. Its migration guide lists several changes that turn working requests into errors:

  • Sampling settings are rejected. Remove temperature, top_p and top_k.
  • Thinking is on by default. A response can now start with a thinking block, so code that reads content[0].text breaks. Read content blocks by their type. This is the bug I fixed in mnemo-cortex #25.
  • Thinking budgets are rejected. Replace budget_tokens with thinking: {"type": "adaptive"} and an output_config.effort level.
  • The old way to turn thinking off fails. {"type": "disabled"} returns a 400 error; send {"type": "between_tools"} instead.
  • Prefill is rejected. The conversation must end with a user message, so move any prefilled assistant text into the system prompt.
  • Forced tool use is rejected. A tool_choice of any or tool returns a 400 error; use auto with strict tools.
  • The same text uses about 30% more tokens. Recheck max_tokens, limits and cost.

3. OpenAI's gpt-5 family

OpenAI's deprecations page lists these replacements:

Shutting down Dec 11 Recommended replacement
gpt-5, o3 gpt-6.1-sol
gpt-5-pro, o3-pro gpt-6.1-sol with reasoning.mode: pro (Responses API)
gpt-5-mini gpt-6.1-terra
gpt-5-nano gpt-6-luna

What to test after you switch:

  • Tool calls. When I moved a Telegram bot from gpt-5-mini to gpt-5.6-terra, the maintainer added one more commit before merging: "disable Terra reasoning when using function tools" (PR #43). If your app uses tools, run those paths before you ship.
  • Reasoning settings. Check that your reasoning effort value is one the new model accepts. GPT-5.6 Terra, for example, doesn't list minimal.
  • Cost. Compare the new model's price per token with what you pay today before you switch everything.

4. Already happened elsewhere

  • Groq shut down llama-3.3-70b-versatile and llama-3.1-8b-instant on Aug 16; calls to them now return errors.
  • xAI redirects eight retired Grok model names, mostly to grok-4.3, which bills at its own rate.
  • Mistral returns a 404 error for retired model IDs.
  • Google lists Nov 17 as the earliest shutdown date for several Gemini preview audio models.

Checklist

  1. Search the whole repo for every form of the old model name.
  2. Remove settings the new model rejects.
  3. Read responses by block type, not by position.
  4. Test tool calls and reasoning settings.
  5. Recheck token limits and cost.

If you maintain an open-source app still on one of these models, tell me and I'll send a free PR. If you'd rather have it done in your private code, I do it for a fixed price, from $400 per repo, paid after you've reviewed it: [your email], or github.com/SamHofmannOfficial.

I used AI tools to help research and write this guide.

Top comments (0)