Meta has released Muse Wonder, a new family of its strongest models for agentic work, in two sizes: Muse Wonder and the larger Muse Wonder Full. By Meta’s published numbers, Full beats Claude Opus 5.5 on 8 of 14 benchmarks, leading 57.5 vs 46.3 on the long-horizon engineering test SWE-Marathon, 79.0 vs 74.2 on DeepSWE 1.1, and 13.1 vs 4.2 on the Harvey legal benchmark, which only counts tasks completed without a hallucination. The models are built around Active Memory and Active Focus and were trained in Meta’s gigawatt Prometheus data center; weights stay closed, with Muse Wonder served through the Meta Model API and Muse Code, and Full limited to Muse Code subscribers. Meta says the runs came from independent third parties, but Full ran at maximum reasoning effort while competitors ran inside Codex and Claude Code, so a strictly controlled comparison is still missing.

Related: Muse Glimmer: Open-Weight 30B Model for Local Agent Workflows, Anthropic Releases Claude Opus 5.5

Meta Muse Wonder