Meta has introduced two new AI models, Muse Wonder and Muse Wonder Full, designed to compete with top offerings from OpenAI and Anthropic. The models are currently available to Muse Code subscribers, with the standard Wonder version also accessible via the Meta Model API.
In the Personal Superintelligence Index, Muse Wonder Full achieved 59.1 points, outperforming GPT-6 Astra at 54.5, Claude Opus 5.5 at 55.2, and Claude Fable 5.1 at 49.9. The standard Muse Wonder model scored 54.3 points in the same assessment.
The models demonstrate strong performance in software development tasks. In the DeepSWE 1.1 benchmark, Muse Wonder Full scored 79 points, while GPT-6 Astra and Claude Opus 5.5 reached 74.1 and 74.2 respectively. In the SWE-Marathon test, the model achieved 57.5 points.
Performance varies across different benchmarks. In Terminal-Bench 4.0, Muse Wonder Full scored 57.1 points, trailing behind GPT-6 Astra and Claude Opus 5.5. In ProgramBench, the model scored 59 points, while Claude Opus 5.5 reached 70.5 points.
Muse Wonder Full showed significant results in the Harvey LAB-AA legal task benchmark, scoring 13.1 points compared to 8.6 for GPT-6 Astra, 6.4 for Claude Fable 5.1, and 4.2 for Claude Opus 5.5. Meta plans to integrate the Wonder architecture into its Muse AI agent.
