CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
← Back to feed

Inside Meta's Muse: Researcher Finds OpenAI and Anthropic Models Shipped in the Runtime

Meta’s Muse AI agent, which the company markets as powered by its own internal model Avocado, ships a runtime that can address models from Anthropic and OpenAI — and at least one session captured in the wild was routed to something labeled azure/muse-special bearing unmistakable OpenAI fingerprints.

The discovery comes from a developer identified as mouse.dev who published a forensic breakdown of Muse’s agent daemon after an earlier article about the runtime’s filesystem attracted attention on Hacker News. In the session logs from their own VM, all but one agent session ran on Avocado. A single subagent on September 21 used azure/muse-special.

The code behind that alias reads: “GPT Responses model client via MAGI native Azure OpenAI lane.” Log lines from the session carry a gpt_responses_v1 signature with an encrypted payload beginning with gAAAAA — a prefix OpenAI uses in its Responses API. Tool call IDs in the session followed the OpenAI format: call_ plus 24 mixed-case alphanumeric characters. All Avocado sessions use call_ plus 32 hex characters. The difference is not subtle.

The full model catalogue shipped inside Muse’s hatch daemon lists roughly 15 Avocado variants and, separately:

  • Claude Opus 4.6, 4.7, and 4.8
  • Claude Sonnet 4.6 and Haiku 4.5
  • GPT-5.5 and GPT-5.6 via direct OpenAI, Azure, and Codex routes
  • Kimi K3 through Fireworks and Meta-hosted paths

The catalogue is a list of addressable model IDs, not a list of models in active use. Muse’s inference-proxy service holds API keys for Anthropic and OpenAI with access restricted to that service.

The researcher stops short of calling this distillation — shipping API access is not the same as training on outputs — but raises the question explicitly. Meta has not commented. The more immediate question is operational: if Muse can route production subagents to GPT or Claude in fallback or specialized scenarios, the clean story that Avocado powers Muse becomes harder to tell.

The broader model catalogue also underscores how deeply frontier agents have come to depend on their competitors’ infrastructure for redundancy, specialization, or cost reasons that no company is eager to advertise.