OpenAI gpt-5.6-luna is incorrectly marked temperature = false

models/openai/gpt-5.6-luna.toml currently declares temperature = false, but the direct OpenAI Chat Completions API now accepts and appears to apply non-default temperature values for Luna when reasoning is disabled.

Direct API verification

Tested against POST https://api.openai.com/v1/chat/completions on 9 September 2026, directly rather than through a proxy or client library.

The minimal request used:

{
  "model": "gpt-5.6-luna",
  "messages": [{ "role": "user", "content": "Reply with OK." }],
  "reasoning_effort": "none",
  "max_completion_tokens": 16,
  "temperature": 0.2
}

This returned 200 and OK. A second request with temperature: 1.5 also returned 200.

I also ran a small sampling check using the same eight-choice prompt:

• Eight requests at 0.2: orchard 6, comet 2.
• Eight requests at 1.5: comet 4, cedar 2, lantern 1, orchard 1.
• Across six paired requests using the same seeds, seed 1 changed from comet at 0.2 to orchard at 1.5; the other five pairs matched.

The sample is too small to characterize Luna's temperature response precisely, but acceptance of both values shows that blanket temperature = false is no longer accurate for this tested configuration. The distribution and paired-seed result are additional evidence that the value is applied rather than merely ignored.

Public corroboration

The generative-ui-bench project sends Luna temperature: 0.7 with reasoning_effort: "none" through direct Chat Completions and has committed 184 Luna outputs:

• Request parameters: https://github.com/thesysdev/generative-ui-bench/blob/07db85f2d9c04d218e0bf1300b8ec96153b8a114/run.ts#L129-L140
• Raw Luna outputs: https://github.com/thesysdev/generative-ui-bench/tree/07db85f2d9c04d218e0bf1300b8ec96153b8a114/raw/luna

Related history

• #5458 and #5459 corrected temperature metadata for several other GPT-5 models, but did not update the GPT-5.6 family.
• The Luna entry still has last_updated = "2026-07-09": https://github.com/anomalyco/models.dev/blob/9fc693ffb14f4ac26d1f21594baf938b05a57ad3/models/openai/gpt-5.6-luna.toml

Suggested change

Please update Luna's temperature metadata to reflect the current API behaviour. If temperature is intended to mean support across every endpoint and reasoning effort, a boolean may be too broad: the evidence here is specifically for Chat Completions with reasoning_effort: "none".

Start with models/openai/gpt-5.6-luna.toml, then compare the temperature metadata and schema used by nearby GPT-5 entries and the changes referenced in #5458 and #5459. Confirm how the field represents endpoint or reasoning-mode support, update Luna’s metadata accordingly, and verify the resulting model definition is consistent with the repository’s conventions.