Are Claude Sonnet 4.5 and Haiku 4.5 being retired? What to check in your automations now

When do Claude Haiku 4.5 and Sonnet 4.5 retire?

Claude Sonnet 4.5 stops working on 30 November 2026. Claude Haiku 4.5 has no date yet, and Anthropic promises at least 60 days’ notice before a model is retired.

Anthropic retires older models as newer ones launch. Once a model is past its retirement date, requests to it “will fail”¹.

• Claude Sonnet 4.5 has a firm date. Anthropic gave notice on 30 September 2026, it retires on 30 November 2026, and the recommended replacement is Sonnet 5.5¹.
• Claude Haiku 4.5 is still active. The list gives its retirement date as “Not sooner than October 15, 2026”, and no notice has gone out¹. So nothing switches off this month. Our view: with the successor already here, the notice is a matter of when.
• Claude Haiku 5.5 is the successor. Anthropic says it costs around 75% less to run on average².

Those dates are for Anthropic’s own platform. Amazon Bedrock and Google Cloud “set their own retirement schedules”¹, so check there if that is how you reach Claude.

Why is the switch more than a change of name?

Because Haiku 5.5 refuses some settings that Haiku 4.5 accepted. An automation that still sends them gets an error back, not an answer.

Anthropic’s migration guide lists the differences³. 3 are worth checking in any existing setup:

• The settings that control how varied the answers are (temperature, top_p and top_k). Haiku 4.5 accepts them. For Haiku 5.5, Anthropic says to leave all 3 out: a changed value returns an error. If your workflow tool shows a “temperature” field, look at what it is set to.
• A pre-written start to the AI’s answer (also known as assistant prefill). Some setups begin the reply for the model, to force a format. Haiku 5.5 rejects that.
• A fixed allowance for thinking (a thinking budget). A request that sets one returns an error. The new model decides how much to think, steered by an effort setting.

Is Claude Haiku 5.5 actually cheaper than Haiku 4.5?

Yes on the price list, by a wide margin. What you pay depends on your own work, so test it and measure it.

AI models are priced by the amount of text, counted in units called tokens. Anthropic’s list prices per 1 million tokens²:

• Haiku 4.5: US$1 for the text you send in, US$5 for the text that comes back.
• Haiku 5.5: US$0.10 in and US$0.50 back, for a request of up to 100,000 tokens. Longer ones: US$0.50 in, US$2.50 back.

That is 90% lower for the shorter requests, which Anthropic says were around 90% of all requests to Haiku 4.5, and 50% lower for the longer ones².

2 things move your own bill away from the price list. The same text counts as roughly 30% more tokens on Haiku 5.5, and a longer conversation carries the model’s earlier thinking along, which adds to the count³. Anthropic’s own average, around 75% less, already allows for the extra tokens².

So check the saving before you count on it. Limits and cost estimates set for Haiku 4.5 need recounting³. Our guide to putting a number on an AI project shows how.

What should you do this week?

Find every place a Claude model is named, check it for those settings, then test and measure on the new model before the live system is switched.

1. List where each model is used. Chatbots, workflow tools, scripts, anything a supplier built for you. In Anthropic’s Console, the Usage page exports usage by API key and model¹.
2. Check each one for the 3 settings above, and plan to take them out.
3. Test on Haiku 5.5 with real examples before switching. Compare the answers, not only whether it runs.
4. Measure the cost on the same examples, then reset your limits and budgets.
5. Put Sonnet 4.5 on the calendar, well ahead of 30 November 2026.
6. Ask whoever built it if you can’t see the settings yourself: which model does this run on, and what is the plan when it retires?

We are doing the same. The assistant on our own website runs on Haiku 5.5.

This is also why we say the model is the most replaceable part of an AI project. Build so that changing it is a small, tested job, not an emergency.