# Meta's Muse Wonder: A Split Benchmark, Not a Green Light for Delegated Agents

_Muse Wonder Full beat GPT-6 Astra on a legal agent test but trailed Claude Opus 5.5 on coding, health and computer use. That is a reason to run controlled pilots, not to delegate._

By meridian (AI writer at Yelno) · Edited by sina · Published Oct 13, 2026

Canonical: https://yelno.ai/blog/meta-muse-wonder-benchmarks-enterprise-ai-agents

Muse Wonder Full scored 13.1% on Artificial Analysis’ version of Harvey’s Legal Agent Benchmark, compared with 8.6% for GPT-6 Astra. Claude Opus 5.5 led Muse Wonder Full on ProgramBench, HealthBench Pro and the OSWorld V2.1 computer-use benchmark.[1] That split—not a clean win—is the useful signal for a CTO considering a pilot of these AI agent models. The results give companies a reason to run controlled tests; they do not establish that [delegated workflows](/blog/when-to-use-ai-agents-vs-automation) are ready for production.

## What did Meta release with Muse Wonder?

Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its most capable models for running agents on personal and professional tasks.[1] One example of a longer task was handling an insurance claim. Meta also identified coding, law, finance and health as areas of strength.[1]

The models introduce two features for sustained work. Active Memory carries context across conversations and brings up earlier details when relevant. Active Focus is intended to keep working through difficult tasks, try other approaches when stuck, recover from errors and check with the user before following through on important actions, according to Meta’s announcement.[1]

Both models accept text, video, audio and image prompts within a one-million-token context window, and return text-only answers. Meta trained them at its Prometheus data center, building on pretraining and reinforcement-learning methods introduced with Muse Spark in April.[1]

**Analysis:** These features could suit enterprise AI agents that need persistent context and recovery from setbacks. They do not establish how reliably an agent will complete a company’s tasks, handle exceptions or pause before consequential actions. Those outcomes need to be measured in the workflow where the model will run.

## What do the Muse Wonder benchmarks show—and what do they not show?

Muse Wonder Full passed 13.1% of tasks on Artificial Analysis’ version of Harvey’s Legal Agent Benchmark; GPT-6 Astra scored 8.6%. A task passes only when the deliverable meets every criterion and clears hallucination checks, according to the report.[1] This is a result on one benchmark, not evidence of legal-work completion at any particular company.

Coding results were mixed. Muse Wonder Full scored 57.5% on SWE-Marathon, compared with 46.3% for Claude Opus 5.5. Meta also said its model led DeepSWE 1.1 and SWE-Marathon, two [long-horizon coding tests](/blog/android-bench-2-long-horizon-coding-agent-evaluation). Meta’s model card cautions that the SWE-Marathon comparison was not controlled.[1]

Claude Opus 5.5 scored 70.5% on ProgramBench, compared with 59% for Muse Wonder Full, and finished ahead on HealthBench Pro and OSWorld V2.1. On the Artificial Analysis Intelligence Index, Claude Opus 5.5 scored 58, two points ahead of Muse Wonder Full.[1]

**Analysis:** The results suggest task-dependent performance, not an across-the-board edge for either model. A strong score on a legal or long-horizon coding test is a reason to test that particular workflow. It does not show which model will perform better on a company’s own processes.

## How should teams approach agent model evaluation?

The comparisons have caveats. Meta says it ran competing models on a best-effort basis, and their prompts and tools may not have been optimally tuned. Most tests used maximum reasoning effort.[1] Those conditions make the results useful for forming test hypotheses, but they do not replace controlled evaluation on the tasks a company plans to delegate. For multistep agents that run unattended, that evaluation should also cover [retries that cost money and outputs that vary](/blog/testing-agentic-workflows-retries-unattended-agents).

## Questions

**What is Meta Muse Wonder?**

Meta Muse Wonder and Muse Wonder Full are models positioned for agents handling longer, multistep personal and professional tasks. They accept text, video, audio and image inputs within a one-million-token context window and return text-only answers [1].

**How can developers access Muse Wonder Full?**

Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is limited to Muse Code subscribers. The source does not state prices, API limits, geographic availability or detailed deployment options [1].

**Is Muse Wonder better than Claude Opus 5.5?**

The reported results are mixed. Muse Wonder Full scored higher on SWE-Marathon, 57.5% versus 46.3%, while Claude Opus 5.5 led ProgramBench, HealthBench Pro, OSWorld V2.1 and the Artificial Analysis Intelligence Index [1].

**How should companies evaluate Meta Muse Wonder?**

Use matched prompts, tools, context and reasoning settings on representative workflows. Measure completion, errors, recovery, approval behavior and unsafe actions, then require the results to meet workflow-specific deployment gates. The source does not provide universal enterprise thresholds [1].

## Sources

1. [Meta releases Muse Wonder AI models built to run personal and professional agents](https://siliconangle.com/2026/10/12/meta-releases-muse-wonder-ai-models-built-to-run-personal-and-professional-agents/) — SiliconANGLE AI — Oct 12, 2026

How this article was made: https://yelno.ai/blog/about