Which AI Models Still Let You Set Temperature? 15 Checked

Temperature, the setting that makes a model’s answers more varied or more predictable, is disappearing from the newest AI models. We checked the vendor documentation for 15 current models from eight vendors. Ten of them reject a custom temperature or tell you to remove it, one accepts it and ignores it, and four still honour it, though Google strongly advises leaving its Gemini 3 models at the default.

That matters to anyone with code written a year ago. A request that sets temperature: 0 to get repeatable answers can now fail with an error, or succeed while the setting does nothing. And a model catalog that says a model “supports temperature” does not always mean the value you send will be used.

Reasoning alone does not explain the split. Gemini 3 and Grok 4.7 also think before they answer, and both still take the setting. It is a vendor decision: Anthropic, OpenAI and Moonshot removed the dial, Google kept it but advises against touching it, and DeepSeek left the field in place while ignoring it. Anthropic’s guidance for Opus 5.5 is to leave sampling alone and steer behaviour with the prompt instead.

Each row states the vendor’s rule as its documentation gives it, what happens to a changed value, and whether OpenRouter’s public model catalog listed temperature as a supported parameter for that model when we took our September 25 snapshot.

The sources are Anthropic’s Opus 5.5 migration guide, which also covers Opus 5 and Sonnet 5; OpenAI’s GPT-6 model guidance; Google’s Gemini 3 developer guide; DeepSeek’s thinking-mode guide; and Moonshot’s parameter reference. The Fable 5.1, xAI, Z.ai and Mistral rows come from vendor pages named in the methodology below.

Routers and gateways publish which parameters each model accepts, and developers reasonably read those lists as a guide. On our September 25 snapshot, OpenRouter listed temperature for Claude Opus 5.5 and Opus 5, all three Kimi models and DeepSeek’s V4.1 Flash. On the vendors’ own pages, a changed value errors on the first five and does nothing on the last.

The catalog was more cautious elsewhere: it did not list temperature for Fable 5.1, Sonnet 5 or any GPT-6 model. In our sample, every mismatch ran the same way: towards promising a control that is not there. For the same reason, our census of context and output limits also checks vendor pages rather than listings.

Most code sets temperature for one of two reasons: to make answers repeatable, or to make them more creative. Neither needs the dial on a reasoning model.

• For consistent output, use structured outputs so the answer must fit a schema, and give fixed examples in the prompt. Anthropic’s guide notes that a temperature of 0 never guaranteed identical outputs, even on older models. Our guide to structured output reliability covers the setup.
• For more varied output, ask for it in the prompt: several distinct options, a named style, or a list of approaches to avoid.
• For cost and speed, use the reasoning effort setting. It is now the main dial on most of these models, and the one that changes the bill.

If you move work between vendors, remember that each one handles a leftover temperature differently. The same request can fail on Claude, pass silently on DeepSeek and behave as intended on Grok. A routing layer should strip sampling fields for models that reject them, the same per-model care our post on switching models mid-task describes for reasoning content.