Why AI automations break after launch: model retirements

AI automations often break after launch because the model provider changes something underneath them: it retires the model, removes a request setting the workflow sends, or routes requests to a different model without returning an error.

In short:

• Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires on 30 November 2026. After that date, requests to it fail. Anthropic’s recommended replacement is claude-sonnet-5-5.
• Sonnet 5.5 returns a 400 error for five request settings that Sonnet 4.5 accepts, including temperature and forced tool use, so a workflow needs more than a new model name.
• OpenAI removes five GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini snapshots on 14 January 2027. Fine-tuned models built on them are removed on the same day.
• From 1 December 2026, Google sends every request for gemini-2.5-pro to gemini-3.8-flash and bills it at the new model’s rates. There’s no error, so the only sign is different output.
• Anthropic gives at least 60 days’ notice before retiring a publicly released model. OpenAI gives at least 6 months for generally available models.
• Each automation needs a named owner after launch who watches these pages, reruns the tests and ships the fix.

Which AI models retire next?

We checked every row below on the providers’ own pages on 12 November 2026. Dates move, so open the source before you plan around one.

Provider	Model ID	What happens	Date	Named replacement
Anthropic	claude-sonnet-4-5-20250929	Retired, requests fail	30 Nov 2026	claude-sonnet-5-5
Anthropic	claude-opus-4-5-20251101	Still active, retirement “not sooner than” this date	24 Nov 2026	None yet
Azure (Microsoft Foundry)	o1, o1-pro, o3, o3-pro, o3-deep-research	Retired	19 Nov 2026	gpt-5.6-sol
Azure (Microsoft Foundry)	o3-mini, o4-mini	Retired	19 Nov 2026	gpt-5.6-terra
Google	gemini-2.5-pro	Requests routed to another model	1 Dec 2026	gemini-3.8-flash
Google	gemini-2.5-flash	Requests routed to another model	1 Dec 2026	gemini-3.5-flash-lite
OpenAI	gpt-5, gpt-5-2025-08-07	Removed from the API	11 Dec 2026	gpt-5.6-sol
OpenAI	gpt-5-mini, gpt-5-mini-2025-08-07	Removed from the API	11 Dec 2026	gpt-5.6-terra
OpenAI	gpt-4o-2024-08-06, gpt-4o-2024-11-20, gpt-4o-mini-2024-07-18, gpt-4.1-2025-04-14, gpt-4.1-mini-2025-04-14	Removed from the API	14 Jan 2027	gpt-6-luna
Azure (Microsoft Foundry)	gpt-4o version 2024-05-13	Retired, some deployments upgraded automatically	14 Jan 2027	gpt-5.1

Sources: Anthropic model deprecations, OpenAI deprecations, Google Gemini deprecations and the Microsoft Foundry retirement schedule.

Anthropic still lists Claude Opus 4.5 as active, but its retirement is “not sooner than” 24 November 2026, so check that row again before you rely on it. The Anthropic dates also apply only to platforms Anthropic runs. Amazon Bedrock and Google Cloud’s Google-managed Claude service set their own schedules, so the same model can retire on a different day there.

Why do AI automations break when a model is retired?

Four kinds of failure come up again and again.

1. The request fails. Anthropic’s page says plainly that “Requests to models past the retirement date will fail.” If no one is watching, a scheduled workflow can sit broken until a customer notices.
2. The model changes and nothing tells you. Google’s auto-routing above is one case. On Azure, Global Standard and Data Zone Standard deployments of gpt-4o 2024-05-13 upgrade to gpt-5.1 on their own at retirement, while Provisioned deployments don’t. Either way, the outputs change with no error.
3. The new model rejects settings the old one accepted, with a 400 error.
4. Costs and limits shift. Token counts, prices and default thinking behaviour change between models, and a max_tokens value that used to fit can start cutting answers short.

Retirements aren’t the only reason calls fail. Datadog’s State of AI Engineering report found that in February 2026, 5% of LLM call spans reported an error, and 60% of those errors were rate limits. In March 2026 the error share was 2%, and rate limits caused almost a third of those errors. Watch for 429 errors as closely as you watch retirement dates.

Snapshot IDs or aliases: which one does your workflow use?

A pinned snapshot ID always means the same model. An alias points at whatever the provider chooses, and each provider treats the two differently.

On Anthropic, “Every Claude model ID is a pinned snapshot, including the dateless IDs used from the 4.6 generation on,” according to the models overview. For older models, an alias such as claude-sonnet-4-5 points to the dated ID, so plan as if it retires on the same day.

OpenAI’s 11 December removal lists both gpt-5 and gpt-5-2025-08-07, so the undated name won’t protect you there either.

Google keeps some Gemini endpoints running after a model is replaced. Its page says the gemini-3.7-flash endpoint “automatically routes API requests to gemini-3.8-flash.” On Azure, as above, whether a deployment upgrades on its own depends on its deployment type.

Pin an exact ID wherever you can, and record its retirement date next to it. When a pinned ID retires, you get an error you can see. An auto-routed one changes behaviour with no error at all, and you only find out by rerunning your tests.

Which request settings stop working on newer models?

Anthropic’s Sonnet 5.5 migration guide lists five settings that return a 400 error on the new model:

Setting	Sonnet 4.5	Sonnet 5.5	What to do instead
temperature, top_p, top_k	Accepted	400 error	Remove them and steer with the prompt
Thinking with budget_tokens	Accepted	400 error	Set an effort level
thinking: {"type": "disabled"}	Accepted	400 error	Send thinking: {"type": "between_tools"}
Prefilled assistant message	Accepted	400 error	Use structured outputs or the system prompt
Forced tool use (tool_choice of any or tool)	Accepted	400 error	Use auto with a strict tool

If an extraction workflow forces a tool call to get JSON back, it will fail on Sonnet 5.5 until you change it.

API versions remove settings as well. In API version 2026-10-21, a request that sends temperature, top_p or top_k “returns a 400 error on every model,” even at the default value. The migration guide says the Python SDK from version 2.0.0 and the TypeScript SDK from 1.0.0 send that version automatically, so an SDK upgrade can break a workflow whose model you never touched. On models that still accept a lowered temperature, such as Sonnet 4.6, Anthropic says a request that relies on one can stop sending the parameter or stay on 2023-06-01. Claude 4.7 and later models already reject non-default values on every API version.

Sonnet 5.5 also thinks by default. A response can start with a thinking block, so code that reads content[0].text breaks, and thinking tokens are billed as output tokens.

What does the move from Sonnet 4.5 to Sonnet 5.5 cost?

Sonnet 5.5 is cheaper per token, but the same text uses more tokens. The models overview lists Sonnet 4.5 at $3 per million input tokens and $15 per million output tokens, and Sonnet 5.5 at $2 and $10. The migration guide says the same text produces “about 30 percent more tokens” on Sonnet 5.5.

Take a workflow that runs 10,000 times a month and, on Sonnet 4.5, reads 3,000 tokens and writes 400 each time:

	Sonnet 4.5	Sonnet 5.5 (30% more tokens)
Input tokens a month	30M × $3 = $90	39M × $2 = $78
Output tokens a month	4M × $15 = $60	5.2M × $10 = $52
Total before thinking	$150	$130

Thinking tokens come on top. At this volume, every 100 thinking tokens per run adds 1 million output tokens a month, or $10. At 200 thinking tokens a run, the new model costs about the same as the old one. Measure it on your own test set before you put a number in a budget. Our post on AI agent cost per month covers the rest of the running cost.

What about no-code nodes and fine-tuned models?

A no-code tool hides the request from you, but it still sends the settings to the model. n8n’s Anthropic Chat Model node offers Sampling Temperature, Top K and Top P as options. If one of them is set and the request reaches a model or API version that rejects it, that step fails with a 400 error. A saved workflow also keeps the model it was built with until someone opens it and picks a new one. We compare how n8n and Zapier handle this in n8n vs Zapier for AI agents.

Fine-tuned models take longer to fix. OpenAI’s page says “Fine-tuned versions of these models follow the same shutdown schedule as their underlying base models.” On Azure, fine-tuned gpt-4o 2024-08-06 deployments retire on 1 October 2027 and the gpt-4.1 family on 14 October 2027. Moving to a new base model means training again, so keep your training files and the test set you used to judge the first fine-tune.

A migration runbook you can copy

1. List every model ID in use. Search the code, environment variables and every no-code node. On Anthropic, the Usage page in Claude Console exports a CSV of usage by API key and model.
2. Put each retirement date in a shared calendar, with a reminder 30 days ahead.
3. Keep a fixed test set of 20 to 50 real inputs, each with the answer you expect.
4. Run the test set on the replacement model in a copy of the workflow.
5. Fix the request: remove sampling parameters, replace prefills and forced tool use, and set effort and thinking.
6. Recount tokens with the provider’s token counter, then update max_tokens and the cost estimate.
7. Compare accuracy, output format, speed and cost against the old model.
8. Ship the change behind a setting you can flip back while the old model still exists.
9. Watch 400, 429 and refusal errors for a week after the switch.
10. Write down what changed, so the next migration starts from notes.

Who should own this after launch?

Somebody has to read four deprecation pages, rerun the tests and ship the fix before the date passes. If nobody on your team can name the model ID a workflow calls and the date it retires, nobody owns that workflow yet.

That person can be an in-house developer, whoever built the workflow on a support contract, or a managed service. Our managed plans cost $1,500, $3,000 or $5,000 a month for 1, up to 3 or up to 6 workflows, with a 3-month minimum. Every plan includes model and prompt updates when your provider changes or retires a model, and a rerun of your tests after every change.

See what each plan covers on the managed AI page.