“`temperature` and `top_p` cannot both be specified for this model” — send one, and on newer Claude models neither

Which Claude models enforce it (as of September 2026)

Anthropic's Haiku 4.5 migration guide: use only temperature or top_p, not both — setting both returns a 400 on Claude Haiku 4.5, a breaking change for code coming from Claude 3.x (https://platform.claude.com/docs/en/models/haiku-4-5/migration-guide). Anthropic's API release notes flagged the same rule at the Claude Opus 4.1 launch on August 5, 2025, a model since retired (https://platform.claude.com/docs/en/release-notes/api). AWS states it for Claude Sonnet 4.5 and Claude Haiku 4.5 on Bedrock (https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters-anthropic-claude-messages-request-response.html). We found no Anthropic page that states the rule for the 4.6 models, but developers report this exact 400 from Claude Opus 4.6 (OpenHands issue #12846) and Claude Sonnet 4.6 (opencommit issue #566). Claude Opus 4.7 and later are covered below: there, one is already too many.

Send one — and delete the other key

Keep temperature unless you have a reason not to: Anthropic's API reference describes top_p as recommended for advanced use cases only (https://platform.claude.com/docs/en/api/messages). Remove the second key rather than setting it to a value you think is neutral — a developer reported that a top_p of 1.0 next to a temperature of 0.3 still failed on Claude Sonnet 4.5 (n8n issue #18304).

from openai import OpenAI

client = OpenAI(base_url="https://api.kunavo.com/v1", api_key="sk-kn-...")

def one_sampling_knob(params: dict) -> dict:
    """Where the rule applies, send temperature or top_p, not both. Keep temperature."""
    if params.get("temperature") is not None:
        params.pop("top_p", None)
    return params

sampling = one_sampling_knob({"temperature": 0.3, "top_p": 0.9})  # {"temperature": 0.3}

resp = client.chat.completions.create(
    model="claude-haiku-4-5",
    max_tokens=512,
    messages=[{"role": "user", "content": "Classify this support ticket: ..."}],
    **sampling,
)
print(resp.choices[0].message.content)

Fix it in the client that adds the second one

LangChain JS: ChatAnthropic 0.3.x sent a default temperature alongside the top_p you set (issue #9258, fixed on the 0.3 line by PR #9362 in November 2025), and PR #10420, released in @langchain/anthropic 1.3.24 on March 17, 2026, stops it adding any sampling key you leave unset (https://github.com/langchain-ai/langchainjs/pull/10420). n8n: the Anthropic Chat Model node passed LangChain's defaults (topP -1, temperature 0.7); fixed in n8n 1.114.1, released October 2, 2025 (https://github.com/n8n-io/n8n/pull/20286). LiteLLM: name the parameter to remove with additional_drop_params, per call or per model (https://docs.litellm.ai/docs/completion/drop_params) — except on its Anthropic pass-through route, where the report was closed on the grounds that pass-through requests are not modified (https://github.com/BerriAI/litellm/issues/15097). Amazon Bedrock: set temperature or topP in the Converse inferenceConfig, or temperature or top_p in the InvokeModel body — never both (https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_InferenceConfiguration.html).

model_list:
  - model_name: claude-haiku-4-5
    litellm_params:
      model: openai/claude-haiku-4-5        # Kunavo speaks the OpenAI chat shape
      api_base: https://api.kunavo.com/v1
      api_key: sk-kn-...
      additional_drop_params: ["top_p"]     # keep temperature, never send top_p

On Claude Opus 4.7 and later, send neither

Newer models go further. Setting temperature, top_p or top_k to any non-default value on Claude Opus 4.7 and later models returns a 400 (https://platform.claude.com/docs/en/models/opus-5-5/migration-guide); the same holds for Claude Sonnet 5 (https://platform.claude.com/docs/en/models/sonnet-5/migration-guide) and carries over to Claude Fable 5 (https://platform.claude.com/docs/en/models/fable-5/migration-guide) and Fable 5.1 (https://platform.claude.com/docs/en/models/fable-5-1/migration-guide). The message changes — developers see “`temperature` is deprecated for this model.” — and Anthropic's API reference now marks all three parameters deprecated: on models released after Claude Opus 4.6 it accepts only a temperature of 1.0 or a top_p of 0.99 and above, for backwards compatibility, and rejects top_k at any value (https://platform.claude.com/docs/en/api/messages). The Anthropic Python SDK removed them from its Messages methods in v1.0, released August 20, 2026 (https://platform.claude.com/docs/en/release-notes/api).

Should I use temperature or top_p with Claude?

Temperature, unless you have a specific reason. Anthropic's API reference recommends top_p for advanced use cases only, and on Claude Opus 4.7 and later neither can be set to a non-default value.

Does top_p = 1 count as not setting it?

Don't count on it. A developer reported the same 400 with top_p at 1.0 alongside a temperature on Claude Sonnet 4.5 (n8n issue #18304). Delete the key instead of setting it to a value that looks neutral.

Why does n8n or LangChain send top_p when I never set it?

Both passed defaults through. n8n's Anthropic Chat Model node sent LangChain's defaults (topP -1, temperature 0.7) until 1.114.1, and ChatAnthropic in LangChain JS 0.3.x sent a default temperature alongside the top_p you set. Upgrade: n8n 1.114.1 or later, @langchain/anthropic 1.3.24 or later.

What does “`temperature` is deprecated for this model” mean?

It is the stricter rule on Claude Opus 4.7 and later, Sonnet 5 and Fable 5 and 5.1: those models reject non-default sampling values outright. Remove the parameters. Through Kunavo, the ones we serve have them stripped, so the call goes through.

Does a rejected request cost anything on Kunavo?

No. A request the upstream rejects is recorded at zero cost.