Peter James found an OpenAI model labeled muse-special inside Meta's Muse

A researcher found evidence that Meta's Muse agent sometimes runs on an OpenAI model instead of Meta's own Avocado. On 2026-09-25, Peter James, writing at mouse.dev, published his second post inspecting the virtual machine that Muse ships with. In the session logs, one subagent ran on a model labeled azure/muse-special, and the payload carried three markers that match the OpenAI Responses API.

What the logs showed

James recorded every model that each Muse session used inside his VM. Almost every session ran on Avocado, Meta's internal model, with about 15 versions of Avocado listed in the daemon's model catalogue. One session on 21 September was routed to azure/muse-special instead. James pulled the raw transcript for that session and compared it against known formats from other providers.

Three details lined up with OpenAI: the request ID began with call_ followed by 32 hex characters, which is OpenAI's usual style rather than the format Avocado sessions print. The signature was tagged gpt_responses_v1, matching the Responses API. And the encrypted reasoning payload started with gAAAAA, which is the prefix OpenAI uses when it hands back a summary of its reasoning. A string in the binary reads "GPT Responses model client via MAGI native Azure OpenAI lane."

What else is in the runtime

The catalogue shipped with the Muse daemon lists more than Avocado. James reports GPT-5.5 and GPT-5.6 variants via OpenAI, Azure and Codex, plus Kimi K3 through Fireworks and Meta-hosted routes. There is a full Anthropic client too, with request handling, prompt conversion and streaming parsers. API key files for Anthropic and OpenAI are present, with access restricted to an inference-proxy service.

The Avocado and non-Avocado sessions were treated differently in one important way. Avocado transcripts write the reasoning text directly and, per the privacy note James cites, can be used to develop Meta's models unless the user opts out. The muse-special session's raw reasoning was encrypted and, per a string in the binary, cannot be sent to the reinforcement learning completion server. So Meta appears to see the reply, the tool calls, and a short reasoning summary. The full reasoning stays with OpenAI.

The consequence

James is careful about what the evidence supports. He explicitly says there is no indication that Meta copies OpenAI or Anthropic weights, because the raw reasoning stays encrypted at the provider. His conclusion, in his own words: muse-special is possibly a GPT model served through Azure, and the runtime has the clients to route to several providers without asking the user.

Two things follow for anyone picking an AI product to build on. The runtime is the source of truth for which model handled a request, and it can disagree with the marketing. Meta has a business reason to keep every session on its own models, and here one session sat on OpenAI instead. A vendor with an inference-proxy service in front of the model list can also change routing at any point without saying so. If where the prompt goes matters to you, verify against the shipped runtime.

Source

Primary source: Is Meta's Muse secretly running an OpenAI model? by Peter James at mouse.dev. Reader discussion: Hacker News thread.