Pre-release of v3.46.0, for field validation of OpenAI GPT-6 support (Discussion #928). HACS offers it only to users who opted into betas. It contains everything on main as of 58b641e.
OpenAI GPT-6 models work, tools included (#993)
Before this release, every request that used a tool failed on a GPT-6 model with "Function tools with reasoning_effort are not supported ... in /v1/chat/completions". The OpenAI provider now uses OpenAI's Responses API for the models that need it:
- GPT-6:
gpt-6-astra,gpt-6.1-sol,gpt-6.1-terraandgpt-6-luna, now in the model lists. - GPT-5.4 and later.
- The
-proandcodexmodels.
What else changes:
- Other models: every other OpenAI model stays on Chat Completions, and the OpenAI-compatible provider is unchanged.
- Reasoning effort: the Thinking / reasoning setting gains Off, Extra-high and Max for GPT-6. Off is
none, or the lowest level on GPT-6 Astra and GPT-6.1 Sol, which have nonone. - Temperature: the Responses API takes no temperature or top-p, so they are left out on those models.
- No stored requests: requests to OpenAI set
store: falseon both APIs, so OpenAI keeps no copy of them. - Reasoning: a GPT-6 model's reasoning is not sent back with the next turn. After a tool result it reasons afresh, as on Chat Completions.
- Fallback: a GPT-6 conversation still falls back to Ollama, Anthropic or Gemini.
- Recovery: a rejected reasoning effort is retried at the model's default. A response that fails mid-stream (a rate limit or server error), or runs out of output tokens with nothing to say, moves on to the next provider.
Tested on a live install with gpt-6-luna: tool turns, a streamed voice reply and a Critical Action PIN, with no errors and no fallback. Max effort answered a multi-step request in under 50 s.
Also in this pre-release (since 3.45.1):
- Voice satellites that recognize the speaker: they personalize answers and keep a history per voice (#978).
- Sentinel:
- Heating and cooling now explain power alerts instead of flagging them (#922).
- An optional full-text record of each pushed alert (#922).
- An always-on power sensor no longer reads as stale (#982).
- A stranger on camera is no longer hidden in a grouped push (#964).
- LLM triage no longer re-asks about the same standing condition every run.
- Sentinel no longer holds up a Home Assistant restart.
- Russian PIN: a Critical Action PIN can be said in Russian digit words.
- Database:
- Fallback: a stopped model call is no longer retried on the fallback provider (#985).
The full notes are in the [Unreleased] section of CHANGELOG.md.