Claude Sonnet 4.5 retires on 30 November. Your automation will outlive the model it runs on

On 30 November, Anthropic switches off Claude Sonnet 4.5 on its API. Any request that still names it will fail. Anthropic told customers on 30 September, in line with the minimum 60 days of notice its deprecation policy promises. The model it replaces had been on sale for about fourteen months.

That is the normal shelf life now. Claude Sonnet 3.5 was released in June 2024 and retired in October 2025. Claude Sonnet 4 shipped in May 2025 and was gone by June 2026. Other providers move at a similar pace, and some change the price before they change the model: Google's Gemini 4 Argon costs $2/$10 per million tokens until 30 November and $4/$20 from 1 December.

If you are paying for an automation, the useful conclusion is simple. The workflow you are buying will probably run for years. The model inside it will not. Plan for at least one model swap a year, and judge any build on how cheaply it survives one.

What actually breaks when a model is retired

The obvious failure is the loud one. The model name stops working, every call returns an error, and the automation stops. That is annoying but easy to spot.

The quieter failures are worse.

**The code stops matching the API.** Models change and so do the interfaces around them. Anthropic's API version from 21 October 2026 removed the `temperature`, `top_p` and `top_k` settings outright. Send them and you get an error back. A pipeline written a year ago that sets temperature to keep outputs consistent now needs a code change, not just a new model name.

**The output shifts in ways nobody notices.** A newer model is usually better on average. That does not mean it does your task the same way. It might write longer summaries, format dates differently, be more cautious about refunds, or call a tool in a different order. If the next step in the chain expects the old shape, it either breaks or, worse, quietly accepts something slightly wrong.

**The bill moves.** A replacement model can cost more per token, use more tokens for the same answer, or both. Nobody signed off on that change. It just shows up on the invoice.

**The cloud you are on has its own clock.** Anthropic's dates apply to its own platforms. Amazon Bedrock and the Google-managed version on Google Cloud set their own retirement schedules for the same model. If your automation runs through one of those, the date you read in a news story may not be your date.

What a build should have so a swap is boring

You do not need to understand the code to ask for these. You need to know they exist.

**The model name lives in one place.** A setting, not something scattered through twenty files. Changing it should be a one-line edit that a competent developer can make in minutes.

**There is a test set made from your real work.** Fifty to a hundred real past inputs, each with the output you consider correct. When a new model comes out, someone runs the set and compares. Without it, "the new model seems fine" means somebody tried three examples by hand. This is the single most valuable thing to ask for, and most builds do not have it.

**Every run is logged.** Input, output, model name, tokens used, cost. When something starts going wrong two weeks after a swap, the logs show you whether the change caused it. We wrote about this [before](/blog/why-we-log-every-agent-run). It matters most at exactly this moment.

**Somebody is subscribed to the notices.** Providers email the account owner when a model is deprecated. If that account belongs to a freelancer who has moved on, nobody reads the email and you find out on retirement day. Make sure the API account is in your company's name and the notices go to an inbox someone checks.

**Someone owns the swap.** Either it is in your support agreement, with a price, or your own team has what it needs to do it. "We'll sort it out when it comes up" is how automations end up silently broken for a week.

When this should change your decision

For a simple, high-volume task like sorting inbound email or pulling fields from invoices, model churn is a minor running cost. A swap and a test run takes an afternoon. Do not let it put you off.

It matters more for long, multi-step agents where the model makes judgment calls. Those are the builds where a behaviour shift does real damage, and where a proper test set is not optional.

It should also make you wary of anything sold as "fine-tuned for you" on a specific model version. When that version retires, the tuning goes with it. Ask what happens then, and what it costs, before you sign.

What to do this week

Find out which model each of your automations calls. Ask whoever built them, or look in the provider dashboard. Anthropic's console has a usage export broken down by model and API key. If anything is on Claude Sonnet 4.5, it has four weeks left. Then ask the second question: if the model changed tomorrow, how would you know the output was still right? If nobody can answer, that is the gap to close first.