Logs suggest Meta's Muse app routes some sessions to an OpenAI model

A single session that was not Avocado

Muse, Meta's personal AI agent app, spins up virtual machines to do its work, and those VMs keep logs of which model handled each agent session. That gave a developer named Pete, publishing on mouse.dev, an unusual window into the product's internals. In his VM, nearly every session was served by Avocado, the model family Meta built for Muse. One subagent session on September 21 was not: it was routed to a model labeled azure/muse-special. The post, a follow-up to an earlier mouse.dev article that reached Hacker News' front page, digs into what that label means.

OpenAI fingerprints

Several details point toward OpenAI. Searching the code shipped with Muse, Pete found a string describing a "GPT Responses model client via MAGI native Azure OpenAI lane", and the model catalogue lists azure/muse-special directly alongside azure/gpt-5.6-sol. The transcript from the outlier session carries a signature tagged gpt_responses_v1 with an encrypted payload beginning gAAAAA, a format associated with OpenAI, and its tool call IDs use call_ plus 24 mixed-case characters, matching OpenAI's convention rather than Avocado's 32 hexadecimal characters.

Pete's conclusion is hedged but clear: his best guess is that muse-special is an OpenAI model served through Azure. The files do not reveal which GPT version it maps to, or why that particular subagent selected it.

A catalogue full of rivals

The agent daemon ships a model catalogue that goes well beyond Meta's own stack. According to the writeup, it lists roughly 15 Avocado versions plus Claude Opus 4.6, 4.7 and 4.8, Sonnet 4.6, Haiku 4.5, GPT-5.5 and 5.6 variants reached through OpenAI, Azure and Codex, and Kimi K3 via Fireworks and Meta-hosted routes. Pete stresses that a shipped ID means the runtime can address a model, not that it was ever used.

The Anthropic support runs deeper than identifiers: the daemon contains dedicated request-handling, prompt-conversion and stream-parsing code, and the VM holds API key files for Anthropic and OpenAI, with access restricted to an inference-proxy service. A runtime environment variable, described in comments as a live kill switch rather than leftover config, can shut off Anthropic routing.

Why ship all of it? Pete offers two explanations: either rival models simply do certain tasks better and Muse routes to them selectively, or the infrastructure exists to compare responses, tool calls and traces for distillation and reinforcement learning. Either way, he notes, the model behind Muse is ultimately a server-side decision that Meta can change without asking users.

What Meta can and cannot see

The evidence argues against Meta distilling from other frontier labs. Raw reasoning returned by muse-special arrives encrypted, and the daemon stores it only to pass back on the next turn; the binary states that this encrypted reasoning cannot use the RL completion-server override. Meta can see the reply, the tool calls and any short reasoning summary, while the RL server refuses the encrypted blobs, and there is no indication Meta copies OpenAI or Anthropic weights.

Avocado is treated differently: its thinking is written plainly into transcripts and is available for RL use, and the accompanying privacy note says conversations can be used to develop AI at Meta unless the user opts out.

Early access opens the same week

The findings landed as Meta widened the product's reach. According to TechCrunch, Meta opened early access requests for new Muse features announced at its Connect 2026 conference, inviting users to share a specific prompt with Muse to join the list. Rather than randomized testing cohorts, Meta is courting AI enthusiasts, a crowd TechCrunch notes tends to use multiple AI apps and understands the market.

The teased features include a digital avatar users can video chat with, expanded shopping partnerships and connectors, a Mac app that will be able to operate the user's computer to complete tasks, and Muse on Meta's lineup of AI glasses, activated by a wake word.

Why it matters

If muse-special is indeed an OpenAI model, then Meta's flagship consumer agent, positioned around its own AI, has at least one lane running through a direct competitor, and the plumbing suggests it could add more at will. For users, the deeper issue is opacity: which model reads your files, writes your code or handles your shopping is a routing decision made server-side, and this is a rare concrete look at that layer.

It also sketches an asymmetry in privacy. Reasoning from Meta's own models flows into training unless you opt out, while reasoning from third-party models stays encrypted between Muse and its provider. With early access now open, computer control on the Mac app coming and glasses support on the roadmap, the number of sessions, and of routing decisions, is set to grow quickly.