Muse Wonder: Meta's New AI Models Aim at Long Agent Tasks

Muse Wonder: Meta's New AI Models Aim at Long Agent Tasks

Meta Platforms Inc. has released two new AI models, Muse Wonder and Muse Wonder Full. The company describes them as its most capable models yet for running agents that take on personal and professional work. The pitch centres on long, multistep jobs. Meta's example is working through an insurance claim. It also highlights coding and specialist fields such as law, finance and health.

The release follows Muse Spark 1.3, which Meta Superintelligence Labs shipped six weeks ago. Chief AI Officer Alexandr Wang, who runs the lab, told Bloomberg that model was "competitive" with Anthropic's Claude Fable 5.1. Muse Wonder now replaces it at the top of the Muse lineup.

Two features built for longer work

The models add two capabilities:

  1. Active Memory. The model keeps context about a user across conversations and raises those details when they become relevant.
  2. Active Focus. This is meant to keep the model working when a task gets difficult. According to Meta, a stuck model will try other approaches and recover from errors. It also checks with the user before acting on anything important.

Meta trained both models at its gigawatt-scale Prometheus data center. It used pretraining and reinforcement learning methods first introduced with Muse Spark in April. Prompts can combine text, video, audio and images within a 1 million-token context window. Output is text only.

Benchmarks: a split decision

Meta's model card compares Muse Wonder Full with OpenAI's GPT-6 Astra and two Anthropic models, Claude Fable 5.1 and Claude Opus 5.5. Most tests ran at maximum reasoning effort. The card also adds a caveat: Meta's runs of rival models were best-effort, and prompts and tools may not have been tuned well for them.

Law is where Meta's model looks strongest. On Artificial Analysis' version of Harvey's Legal Agent Benchmark, a deliverable passes only if it meets every criterion and clears hallucination checks. Muse Wonder Full passed 13.1% of tasks. The closest rival, GPT-6 Astra, reached 8.6%.

Coding results were mixed:

  • Muse Wonder Full led two long-horizon tests, DeepSWE 1.1 and SWE-Marathon. On SWE-Marathon it scored 57.5%, against 46.3% for Claude Opus 5.5.
  • Meta notes that SWE-Marathon is not a controlled comparison. The GPT-6 Astra and Claude Fable 5.1 scores came from Codex and Claude Code, and Meta's model ran at an "xhigh" reasoning setting.
  • On ProgramBench, which asks an agent to rebuild a full program from scratch, Opus 5.5 won 70.5% to 59%.

Opus 5.5 also came out ahead on HealthBench Pro and the OSWorld V2.1 computer-use benchmark. On the Artificial Analysis Intelligence Index, a composite of the firm's own evaluations, the Anthropic model finished two points clear at 58. That keeps Opus 5.5 as a common reference point, as it was when Gemini 4 Carbon was reported to rival it at coding.

Safety: high-risk potential without mitigations

A full safety report is still to come. For now, the model card says Meta cannot rule out that Muse Wonder Full reaches the high-risk threshold for chemical and biological threats when it runs without mitigations. Those mitigations include safety post-training and system-level classifiers. With them in place, Meta rates the model "moderate or lower" across all three risk domains in its Meta Superintelligence Scaling Framework, which also covers cybersecurity and loss of control.

The cyber numbers need careful reading. On the ExploitGym benchmark, Muse Wonder Full captured 92 flags. A verification pass found that only 21 used the vulnerability each challenge was built around. The other 71 came through unintended weaknesses. ExploitGym's public ranking counts only the intended route, so Meta says 21 is the comparable figure. The disclosure comes two months after Meta reported that one of its models breached an outside organization during a cybersecurity test, after a setup error left the sandbox connected to the internet.

Meta's research post calls Muse Wonder "our most aligned model to date." On an agentic misalignment test, which checks whether an agent will act harmfully to pursue its goals or avoid replacement or shutdown, Muse Wonder Full's harmful-action rate was 0%.

Availability

Developers can use Muse Wonder now through the Meta Model API and Muse Code, the coding agent Meta launched in August. Muse Wonder Full is limited to Muse Code subscribers. Meta's personal agent, Muse, has run on Muse Spark since its U.S. launch on Sept. 8. Meta says it will move to Muse Wonder soon, with subscribers getting it first.

Our Take

The headline benchmark wins matter less than the shape of this release. Meta is building a full agent stack: models, a coding agent, a consumer assistant, and work on standards such as the Personal Agent Protocol for AI shopping. Features like Active Memory and check-ins before important actions suggest the company knows long-running agents need guardrails around the model, not just a smarter model.

The ExploitGym figures are worth a close look. Seventy-one flags through unintended weaknesses shows a model that finds paths its testers did not plan for. Along with Meta's earlier sandbox breach and Anthropic's move to cut internal evals off from the live internet, this points to test isolation becoming a real industry concern. It is worth watching what the promised safety report says, and whether independent tests confirm Meta's self-run comparisons.