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.
- 01Ten of 15 models refuse a custom temperature or say to remove it.All four Claude models checked, all three GPT-6 models (unless reasoning is off) and all three Kimi models.
- 02DeepSeek accepts it and silently ignores it.In thinking mode, which is on by default, the value has no effect and no error is raised.
- 03Only four models honour a changed value.Gemini 3 (with a strong warning to stay at 1.0), Grok 4.7, GLM-5.3 and Mistral’s API models.
- 04A catalog listing is not a guarantee.OpenRouter’s catalog listed temperature for six of the eleven models where a changed value errors or does nothing.
01 — The countMost new models took the dial away #
Changing temperature fails or is disallowed
Anthropic and Moonshot return an error for a non-default value. OpenAI tells you to remove the parameter whenever GPT-6 is reasoning.
Accepted, with no effect
DeepSeek’s thinking mode takes the parameter for compatibility and ignores it.
A changed value is used
Google, xAI, Z.ai and Mistral still accept temperature. Google recommends not changing it on Gemini 3.
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.
02 — The tableModel by model #
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.
| Sources: each vendor’s API documentation, read September 29, 2026; catalog column from OpenRouter’s models API, observed September 25, 2026 at 08:01 UTC. | |||
|---|---|---|---|
| Model | Changed value | Catalog lists it | Vendor rule |
| --- | --- | --- | --- |
| Claude Opus 5.5 | Error | Yes | Default only. Any other temperature, top_p or top_k value returns a 400 error. |
| Claude Opus 5 | Error | Yes | Default only. Anthropic applies the rule to Claude Opus 4.7 and later models. |
| Claude Fable 5.1 | Error | No | Default only. Anthropic says the sampling-parameter restriction carries over from Claude Opus 5. |
| Claude Sonnet 5 | Error | No | Default only. Non-default sampling returns a 400 error. |
| GPT-6 Astra | Unsupported | No | Remove temperature and top_p. Astra has no “none” reasoning effort, so this always applies. |
| GPT-6 Sol | Unsupported | No | Remove temperature and top_p unless reasoning effort is set to “none”. |
| GPT-6 Luna | Unsupported | No | Same as Sol: allowed only with reasoning effort “none”. |
| Kimi K3 | Error | Yes | Fixed at 1.0; other values return an error. |
| Kimi K2.7 Code | Error | Yes | Fixed at 1.0; other values return an error. |
| Kimi K2.6 | Error | Yes | Fixed at 1.0 with thinking, 0.6 without; other values return an error. |
| DeepSeek, thinking mode | Ignored | Yes | Temperature is not supported in thinking mode, which is on by default. No error is raised. |
| Gemini 3 models | Accepted, discouraged | Yes | Accepted. Google strongly recommends the default of 1.0 and warns lower values can cause looping. |
| Grok 4.7 | Accepted | Yes | Accepted. xAI’s reasoning guide restricts penalties and stop sequences, not temperature. |
| GLM-5.3 (Z.ai) | Accepted | Yes | Accepted. Z.ai says the default depends on the model. |
| Mistral API models | Accepted | Yes | Accepted. Mistral recommends values between 0.0 and 0.7. |
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.
03 — The catchWhere the catalog and the vendor disagree #
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.
“Supported” in a catalog means the route will accept the field in a request. It does not tell you whether a non-default value is allowed, used or silently discarded. For any setting your results depend on, the vendor’s documentation is the authority, and a quick test request is the proof.
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.
04 — AlternativesWhat to use instead of temperature #
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 tostructured 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.
05 — MethodologyHow we checked #
One row per model, one claim per row, taken from the vendor’s own API documentation. The catalog column is a separate observation.
- What was collected
- The documented handling of the temperature parameter for 15 current models from Anthropic, OpenAI, Google, DeepSeek, Moonshot, xAI, Z.ai and Mistral, and whether OpenRouter’s models API listed temperature as a supported parameter for the same model.
- Sources
- Anthropic, Opus 5.5 and Fable 5.1 migration guides; OpenAI, GPT-6 model guidance; Google, Gemini 3 developer guide; DeepSeek, thinking-mode guide; Moonshot, model parameter reference; xAI, reasoning guide; Z.ai, parameter concepts page; Mistral, chat completions API reference. Catalog: OpenRouter models API, all output modalities.
- As-of date
- Vendor documentation was read on September 29, 2026. The catalog snapshot was taken on September 25, 2026 at 08:01 UTC. Models released after September 27 are not included.
- Classification
- “Error” means the vendor states a non-default value is rejected. “Unsupported” means the vendor says to remove the parameter without stating the response. “Ignored” means accepted with no effect. “Accepted” means the vendor documents the parameter with no restriction on its value.
- Exclusions
- Vendors whose documentation we could not read with a clear rule, including Alibaba’s Qwen, are left out rather than inferred. Older models still on sale, such as Claude Haiku 4.5, are not covered.
- Known limitations
- We did not send test requests; every row is the vendor’s written rule. DeepSeek’s non-thinking mode is not a separate row, and its top_p is fixed there. Google’s rule applies to all Gemini 3 models and Mistral’s to its whole chat API, so each vendor has one row.
- Refresh
- Refreshed in place when a listed vendor changes its sampling rules or ships a new flagship model.
06 — ConclusionTemperature is now a legacy setting on reasoning models #
Search your codebase for temperature and top_p, and remove them from every request that goes to a reasoning model
A leftover sampling parameter is either an error waiting for the next model upgrade or a setting that silently does nothing. Both are cheap to find. If you run several models behind one interface and want the request layer cleaned up, our AI transformation team audits model integrations as part of migrations. For the other breaking changes in Anthropic’s newest flagship, see our Claude Opus 5.5 launch coverage.