Claude API 2026-10-21: temperature, top_p and top_k are gone

In Claude API version 2026-10-21, a request that sends temperature, top_p or top_k fails with a 400 error, even when the value is the default. Remove the 3 parameters from your requests before you move to that version. Without them, Claude samples at the defaults, a temperature of 1.0, and for Claude 4.7 and every later model nothing about the output changes.

The change reaches most teams through an upgrade rather than a decision. A Claude SDK sends the API version that was current when its major version came out, so the next major upgrade of the Python or TypeScript SDK moves every request to 2026-10-21 at once. Code that sets temperature=0 will stop working on the day the dependency is bumped.

Who it breaks, and who it does not

It depends on 2 things: the API version the request sends, and the model.

• Claude 4.7 and later (including Opus 5.5, Sonnet 5.5 and Haiku 5.5) already accepted only the default values on the old version, 2023-06-01. If your code runs on these models today, it sends no sampling values or sends the defaults. Under 2026-10-21, even the default is an error, so delete the lines and the output stays the same.
• Older models that accepted other values, such as Claude Sonnet 4.6, are where the change bites. A request that set temperature: 0.2 for steadier output works on 2023-06-01 and fails on 2026-10-21. Removing the parameter makes it work again, at temperature 1.0, so the steadiness is gone and Anthropic says plainly that no other setting replaces it.
• Requests with no sampling parameters are not affected by this part at all.

The same rule covers token counting and every request inside a Message Batch created with the new version.

Try your own request:

POST /v1/messages
anthropic-version: 2026-10-21

{
  "model": "claude-sonnet-4-6",
  "temperature": 0.2,
  "max_tokens": 1024,
  "messages": [ { "role": "user", … } ]
}

Why the parameters went

TemperatureTemperatureA setting that controls how much chance goes into each word a model picks. Low gives focused, predictable answers; high gives more varied ones. scales how the model picks each next TokenTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens.: low values favour the likeliest token, high values spread the choice. top_p and top_k cut the list of candidates before the pick. Anthropic first locked all 3 to their defaults on Claude 4.7 and later, then removed them from the request format and from the SDKs' types in the new API version, with one recommendation in their place: guide the model with the prompt.

That is a change in habit: consistency now comes from what you ask and how you check the answer, not from a number in the request.

What to use instead of a low temperature

Teams set temperature low for 3 reasons, and each has a better tool now.

• The output has to fit a format. Use structured outputsStructured OutputA setting that makes a model's answer follow a JSON schema you give it, so your code can read fields such as a category or an order number without parsing free text., so the reply matches a JSON schema every time, or a tool definition the model fills in. A schema guarantees the shape. Temperature only made a wrong shape less likely.
• The same input should give the same answer. Write the rules down. Give the exact categories, a few worked examples and what to do when unsure, in the system promptSystem PromptInstructions the builder of an AI app gives the model before you type anything. It sets the role, the tone and the rules, and it outranks what users say.. Most variation at temperature 1.0 comes from instructions that leave room for it.
• You want to know it still works. Keep a set of real inputs with the answers you expect and run it after every prompt or model change. An eval measures consistency; a low temperature only hoped for it. How to evaluate an AI agent shows how to build one, and our eval sample size calculator says how many questions tell a real change from noise.

If a request on Sonnet 4.6 truly depends on a low temperature and you cannot test a replacement yet, the guide allows one way out: keep that request on 2023-06-01 while the rest of the code moves. Treat it as temporary: Anthropic retires older models on a published schedule, and every model from Claude 4.7 on accepts only the defaults.

The other changes in 2026-10-21

Sampling is the change most people will hit, but the same version has 3 more that return errors or change fields:

The new SDK major versions also send the API key as Authorization: Bearer instead of in the X-Api-Key header. The API accepts both, but a proxy or gateway that reads X-Api-Key will see no key. Check that before you upgrade, not after.

How to move, in order

1. Search your code for temperature, top_p and top_k in every call to Claude, including batch jobs, eval scripts and any wrapper library that sets them for you.
2. Delete them where the model is Claude 4.7 or later. Nothing changes in the output.
3. For older models with a non-default value, replace the setting with a format, examples and an eval, then delete it. Our LLM API module covers the request shape if you are starting fresh.
4. Check the other 3 changes in the table above.
5. Upgrade the SDK to its new major version, or set the anthropic-version header to 2026-10-21 if you call the API directly, and run your tests before you release.