“Unsupported value: 'temperature' does not support … with this model” — reasoning models take effort and verbosity instead

Remove the sampling parameters

It is a deletion, not a new value. Leave temperature and top_p out of reasoning-model calls entirely, together with logprobs and top_logprobs, which the same rule covers. If you want to change how much the model works, reasoning effort is the knob that is left.

from openai import OpenAI

client = OpenAI(base_url="https://api.kunavo.com/v1", api_key="sk-kn-...")
msgs = [{"role": "user", "content": "Summarize this diff as three changelog bullets."}]

# Before — sampling parameters on a reasoning model
resp = client.chat.completions.create(
    model="gpt-6-astra", temperature=0.2, top_p=0.9, messages=msgs)

# After — removed; reasoning effort is the control that remains
resp = client.chat.completions.create(
    model="gpt-6-astra", reasoning_effort="low", messages=msgs)

Check which models accept sampling at all

OpenAI states the rule per model family, and it turns on reasoning effort. This is its model guidance as read on September 23, 2026. The GPT-5.5 and GPT-5.6 pages do not restate the rule, so treat the silence as a reason to test, not as permission; the o-series rows come from error reports rather than documentation.

gpt-5, gpt-5-mini, gpt-5-nano   error, whatever the reasoning setting
GPT-5.1, GPT-5.2, GPT-5.4       accepted only with reasoning effort "none"
GPT-6 Sol, GPT-6 Luna           remove them unless reasoning effort is "none"
GPT-6 Astra                     remove them: Astra has no "none" effort
GPT-5.5, GPT-5.6                not stated on their guidance pages
o1-preview, o3-mini             rejected (reported errors, not docs)

Steer with the controls that replaced it

OpenAI's guidance names three substitutes when reasoning is on: reasoning depth (reasoning.effort, whose values are model-dependent and run from none and minimal up to xhigh and max), output verbosity (text.verbosity: low, medium or high, medium by default) and output length (max_output_tokens, which counts reasoning tokens as well as visible ones). Chat Completions spells the first two reasoning_effort and verbosity. The guidance offers no sampling-level stand-in for temperature 0; if you used it to keep output repeatable, pin the shape instead, with Structured Outputs for a schema and the prompt for the rest.

# client as in fix.py above
resp = client.responses.create(
    model="gpt-5-6-sol",
    input="Summarize this diff as three changelog bullets.",
    reasoning={"effort": "low"},   # how much it thinks
    text={"verbosity": "low"},     # how much it says
    max_output_tokens=2000,        # hard cap, reasoning tokens included
)
print(resp.output_text)

Put the rule in one helper

Branching at every call site is how the next model family turns into another round of edits. Filter the sampling parameters once, keyed on the model and the effort you are about to send; on a model that supports effort none, they pass through when you ask for none.

SAMPLING = ("temperature", "top_p", "logprobs", "top_logprobs")
REASONING = ("gpt-5", "gpt-6", "o1", "o3", "o4")  # extend as you adopt models

def sampling_params(model: str, effort: str | None, **params) -> dict:
    """Keep sampling parameters only where they are accepted: effort "none"."""
    if model.startswith(REASONING) and effort != "none":
        return {k: v for k, v in params.items() if k not in SAMPLING}
    return params

effort = "low"  # client and msgs as in fix.py above
resp = client.chat.completions.create(
    model="gpt-5-6-terra",
    messages=msgs,
    reasoning_effort=effort,
    **sampling_params("gpt-5-6-terra", effort, temperature=0.2, top_p=0.9),
)

Can I set temperature on GPT-5?

Not on gpt-5, gpt-5-mini or gpt-5-nano: OpenAI says requests that include it raise an error. GPT-5.1, 5.2 and 5.4 accept temperature, top_p and logprobs only with reasoning effort set to none, and OpenAI's GPT-6 guidance says to remove them whenever reasoning effort is not none.

What replaces temperature on reasoning models?

Reasoning effort for how much the model thinks, text.verbosity (verbosity in Chat Completions) for how much it says, and max_output_tokens as a hard cap. Through Kunavo, set verbosity on /v1/responses: the chat endpoint does not forward it to GPT models. For repeatable structure, use Structured Outputs and explicit format rules in the prompt.

Should I just send temperature=1?

Omit it. The Unsupported value message says the default of 1 is accepted on the model that produced it, but other models reject the parameter at any value ("Unsupported parameter: 'temperature' is not supported with this model"), and OpenAI's guidance is to remove the fields.

Why do I get this error when my code never sets temperature?

Something in your stack set it. A jupyter-ai user's o1-preview calls carried 0.7, the default of LangChain's ChatOpenAI class underneath jupyter-ai, and a Cline contributor put the 0 a GPT-5 user hit down to a hard-coded default. Check what your SDK or framework actually puts on the wire, then override or remove it per model.

Does Kunavo remove temperature for me?

Not for GPT models: Kunavo forwards temperature and top_p exactly as you send them, on /v1/chat/completions and /v1/responses alike. It does remove temperature, top_p and top_k on the six Claude models Anthropic removed them from.