Meta's Muse Routes Some Sessions to OpenAI via Azure

A developer who inspected the virtual machine behind Meta's Muse found that one of its subagents runs on a model called azure/muse-special, and its technical fingerprints match OpenAI's API conventions rather than Meta's own systems.

Peter James published the finding on his Mouse blog on Sept. 25, after Muse built him a website and he examined the session logs and files inside the virtual machine it runs in. Nearly every session in his logs routed to Meta's own model, which the files label Avocado. One subagent session instead ran on azure/muse-special.

What the logs actually showed

James found that the muse-special session carried a signature tagged gpt_responses_v1, with an encrypted payload beginning gAAAAA, the format OpenAI's Responses API uses. Its tool call identifiers used the prefix call_ followed by 24 mixed-case characters. The Avocado session, by contrast, used a 32-character hexadecimal identifier and routed through Meta's own inference infrastructure.

Session	Identifier format	Reasoning visible to Meta
Avocado (Meta's model)	32-character hex	Yes, logged for reinforcement learning unless a user opts out
muse-special	call_ plus 24 mixed-case characters	No, encrypted at the source

The daemon James examined also listed a broader catalog of models Muse can call, including several Claude models, multiple GPT-5.x variants and Moonshot's Kimi K3, alongside client code for Anthropic's API and a "kill-switch" setting.

Why Meta might reach outside its own model

Meta launched Muse in early September as a personal AI agent that connects to a user's email, calendar and desktop apps, running on a model Meta calls Muse Spark, with Chief AI Officer Alexandr Wang leading the rollout. James wrote that Meta may route individual tasks to an external model when Avocado underperforms, or to run controlled comparisons between providers without asking users to switch anything themselves. James also wrote that Meta had reached out to him about an earlier post he published on Muse, walking him through its reasoning and calling the company's response "amazing."

Muse's model catalog matters beyond Meta. If muse-special is confirmed as an OpenAI model, it would mean two of the largest AI labs in the world are already routing paying customers' agent sessions through each other's infrastructure, a fact neither company volunteered before James found it in a log file.

Muse is not the only Meta product to face outside scrutiny this month. Staff at the company flagged security flaws in Muse shortly after launch, and OpenAI has its own commercial reasons to keep such arrangements quiet: it sells a competing agents product built on the same kind of cloud infrastructure that James says is carrying some of Meta's traffic.

What is still unconfirmed

Neither Meta nor OpenAI has confirmed which OpenAI model sits behind muse-special. It is also not established how often Muse reaches for the external model rather than Avocado, since James's evidence covers one session inside one virtual machine. Meta has not said whether it has a commercial agreement with OpenAI to serve those requests through Azure, or whether the routing happens without one.