AWS Bedrock (Responses) rejects temperature

medium · inference · probed live 2026-08-01 · affects AWS Bedrock (Responses)

temperature is a standard sampling parameter of the Responses dialect; temperature: 0 in particular is what makes runs reproducible. Bedrock's Responses lane rejects the parameter with a 400 unsupported_parameter error — deterministic sampling is unreachable, so every run on this lane is stochastic.

What to send

(download)

{
  "model": "google.gemma-4-31b",
  "input": [
    {
      "role": "user",
      "content": "Reply with just 'ok'."
    }
  ],
  "store": false,
  "temperature": 0,
  "max_output_tokens": 16
}

What comes back

Verbatim:

{
  "error": {
    "code": "unsupported_parameter",
    "message": "Unsupported parameter: 'temperature' is not supported with this model.",
    "param": "temperature",
    "type": "invalid_request_error"
  }
}

The identical body minus the temperature line succeeds — that pair is the repro. The rejection is why this suite's own probes omit temperature on this lane, and why single-run verdicts on it need repetitions.

Why it matters

Evals and regression tests need determinism; applications porting request bodies across Responses-dialect providers must special-case this one or accept stochastic output on it.

Run it yourself

curl "https://bedrock-mantle.us-west-2.api.aws/openai/v1/responses" \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d @temperature.responses.json

Expected: 200 (the parameter is part of the dialect). Received: 400 unsupported_parameter. Delete the temperature line and the same request succeeds.