Meta releases Muse Wonder AI models designed to guide personal and professional agents

Meta Platforms Inc. today released Muse Wonder and Muse Wonder Full, two artificial intelligence models that it says are the most capable to date of running agents that complete personal and professional tasks.

Meta aimed the release at longer, multi-stage jobs that hand people off to an AI agent. An example of this was processing an insurance claim. Coding is a strength, the company says, as are specialty areas such as law, finance and health.

Active Memory, one of two new features in the models, carries a user’s context from one conversation to the next, displaying details as they become relevant. The other, Active Focus, is intended to keep the model running when a task becomes difficult. According to Meta’s announcement, a stuck model will try other approaches and recover from errors, and it will report back to the user before doing anything important.

Meta trained both models in its gigawatt-scale Prometheus data center, building on pre-training and reinforcement learning methods it introduced with Muse Spark in April. Prompts can mix text with video, audio and images in a contextual window with 1 million tokens. Answers only come back in text.

In the model map, Meta Muse compared Wonder Full to the OpenAI Group PBC’s GPT-6 Astra and two Anthropic PBC models, Claude Fable 5.1 and Claude Opus 5.5. In most tests, each model was run with maximum reasoning effort. Meta’s attempts at competing models were best-effort models, and the map alerts and prompts and tools may not have been optimally tailored to them.

Muse Wonder Full returned a passing score on 13.1% of tasks in the Artificial Analysis version of Harvey’s Legal Agent Benchmark, with a score only counting if it meets all criteria and passes hallucination checks. GPT-6 Astra came closest among competitors at 8.6%.

The coding results were divided. Muse Wonder Full topped two long-horizon tests, DeepSWE 1.1 and SWE-Marathon, scoring 57.5% on the latter versus 46.3% on Claude Opus 5.5. The map indicates that the SWE marathon numbers are not a controlled comparison. Codex and Claude Code produced the ranking results for GPT-6 Astra and Claude Fable 5.1, and Meta’s model ran with a reasoning setting of “xhigh” for this test. ProgramBench, which asks an agent to create an entire program from scratch, scored Opus 5.5 70.5% to 59%.

Claude Opus 5.5 also took first place in HealthBench Pro and the OSWorld V2.1 computer usage benchmark. On the Artificial Analysis Intelligence Index, a compilation of assessments conducted by the benchmarking company itself, the Anthropic model scored 58 points, a two-point lead.

Meta has promised to provide a full security report soon. Meanwhile, the model card states that Meta does not rule out Muse Wonder Full reaching the high-risk threshold for chemical and biological threats if it runs without remedial measures. Mitigations targeting this area include security post-training and system-level classifiers. With these protections in place, the model is rated “moderate or lower” in all three risk areas in the Meta Superintelligence Scaling Framework, including cybersecurity and loss of control.

Of the 92 flags that Muse Wonder Full caught on the ExploitGym cyber benchmark, a review run revealed that 21 were caused by the vulnerability around which each challenge was designed. The model got the other 71 due to unintentional weaknesses. Since ExploitGym’s published ranking only takes into account the intended route, 21 is the number to compare with, the company says. The cyber findings come two months after Meta revealed that one of its models had breached an external organization during a cybersecurity test where a setup error resulted in the sandbox being connected to the internet.

Meta’s research paper calls Muse Wonder “our best-tuned model to date.” In an agent misalignment test, which checks whether an agent is acting harmfully to advance its goals or avoid replacement or shutdown, Muse Wonder Full’s harmful action rate was 0%.

Muse Wonder takes over the top of the Muse series from Muse Spark 1.3, which Meta Superintelligence Labs released six weeks ago. Chief AI Officer Alexandr Wang, who leads the lab, described this model to Bloomberg as “competitive” to Claude Fable 5.1. Meta’s personal agent Muse has been running on Muse Spark since its US launch on September 8th.

Developers can use Muse Wonder starting today through the Meta Model API and Muse Code, the Meta coding agent launched in August. Muse Wonder Full is limited to Muse Code subscribers. Meta said Muse Wonder will soon power its Muse Agent, with subscribers getting it first.

Image: Meta