LLM model deprecations this winter: Claude Sonnet 4.5, Gemini 2.5 and GPT-4o

Three LLM model deprecations land in the next ten weeks. On 30 November Anthropic switches off Claude Sonnet 4.5. On 1 December Google starts answering requests for Gemini 2.5 Pro and Flash with newer models. On 14 January OpenAI removes the GPT-4o and GPT-4.1 snapshots. If a feature in your business runs on one of these models, it will either break or quietly change. Here is what happens per provider, and how to test the replacement before the date does it for you.

What changes, and when

Date	Provider	Model	What happens	Replacement
30 Nov 2026	Anthropic	claude-sonnet-4-5-20250929	Requests fail	claude-sonnet-5-5
30 Nov 2026	OpenAI	Agent Builder, Evals, v1/prompts	Products shut down	Agents SDK, prompts in your own code
1 Dec 2026	Google	gemini-2.5-pro	Routed to another model	gemini-3.8-flash
1 Dec 2026	Google	gemini-2.5-flash	Routed to another model	gemini-3.5-flash-lite
14 Jan 2027	OpenAI	gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini snapshots	Removed from the API	gpt-6-luna

All dates come from the deprecation pages of Anthropic, Google and OpenAI. Google calls its dates the earliest possible ones, so check its page again before you plan around them.

Anthropic: Claude Sonnet 4.5 stops answering

Anthropic told users on 30 September that Sonnet 4.5 retires on 30 November. After that date, requests to it fail. Your code gets an error back, not a slower or different answer. That’s the easy kind of change: you notice it at once. Still, you’d rather notice it in a test than in production.

One trap when you move to Sonnet 5.5: the newer models don’t accept temperature, top_p or top_k at non-default values. Anthropic’s page says they return a 400 error on Claude 4.7 and later models, and API version 2026-10-21 removes the parameters for every model. Many extraction pipelines set temperature: 0 to get stable output. Change only the model name and those calls fail as well.

The dates apply to the Claude API and the platforms Anthropic runs. Amazon Bedrock and Google Cloud’s own Claude offering set their own schedules, so check those separately.

Google: Gemini 2.5 keeps answering, with a different model

Google takes another route. From 1 December, requests to gemini-2.5-pro go to gemini-3.8-flash, and requests to gemini-2.5-flash go to gemini-3.5-flash-lite. You pay the rates of the new model. Google’s page asks you to read the breaking API changes in the migration guide for each new model before that date.

This is the riskier kind of change. Nothing errors, so nobody gets paged. But a different model produces different output. An extraction step may start to miss fields or format dates another way. A classifier may shift its labels. A prompt that was tuned for 2.5 Pro may behave differently on a Flash model. You only find out if you measure it.

Note that gemini-2.5-flash-lite is not deprecated and stays available for now.

OpenAI: GPT-4o and GPT-4.1 snapshots in January

On 14 October OpenAI announced that the dated snapshots of GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the API on 14 January 2027. The recommended replacement for all of them is gpt-6-luna. Fine-tuned models built on these snapshots follow the same schedule, so a fine-tune on gpt-4o-mini-2024-07-18 stops working too. You’ll have to train it again on a newer base model or replace it with a prompt.

Two more dates on 30 November matter if you built on OpenAI’s own tooling. Agent Builder shuts down, and so do the Evals dashboard and API and the reusable prompts API (v1/prompts). OpenAI advises moving prompt content into your application code. If your prompts only live in OpenAI’s dashboard, copy them out now.

Find out which models you use

Most companies don’t know this off-hand. The model name sits in a config file, an environment variable, a no-code tool or a supplier’s product. Look in these places:

• Usage exports. In the Claude Console, export your usage as CSV. It’s broken down by API key and by model. OpenAI’s GET /v1/models endpoint has a shutdown_date field per model that you can check from a script.
• Your code and config. Search for sonnet-4-5, gemini-2.5, gpt-4o and gpt-4.1.
• Tools and suppliers. Automation tools such as n8n or Make, and software you buy, often call a model on your behalf. Ask the supplier which model they use and what their plan is.
• The mailbox of the account owner. All three providers send notices by email. Check that the address on each account still reaches someone who reads it.

What to do now

1. List every feature that calls an LLM, with provider, exact model name and owner.
2. Build a small test set per feature: 30 to 100 real inputs with the output you expect. Remove personal data first; our article on using LLMs without sending personal data shows how.
3. Run the test set against the replacement model and compare. Look at what matters for the feature: missing fields, wrong labels, length, cost per request.
4. Fix the prompts and parameters, for example remove temperature for the new Claude models.
5. Switch before the deadline, on a day you choose, with a way to switch back.
6. For Gemini 2.5, decide before 1 December: move yourself, or accept the automatic switch and watch the output closely in the first days.

How we approach this

We put the model name in one place: an API layer between the application and the model providers. For a financial services company we built a private AI workspace with many agents that works this way. The layer logs which model answers each request, so we can see right away which features depend on a model that’s about to retire. Each feature also has its own test set, so a model switch is a measured change instead of a guess. Read more on our page about AI assistants and agents, or in our article on securing an LLM gateway after the LiteLLM attack.