Model Overview & Capabilities
Muse Wonder is Meta’s premier closed-source frontier reasoning model introduced on October 12, 2026, standing at Rank #6 on the Vibe Coder Journal leaderboard with an empirical composite index score of 50.8. Powered by Meta’s advanced Active Memory architecture, Muse Wonder scores 51.5% on SWE-bench Verified, 53.0% on Terminal-Bench 2.0, 48.0% on Aider Polyglot, and 88.5% on GPQA Diamond, establishing Meta’s strongest commercial software engineering system to date.
The model is billed at $4.00 per million input tokens and $20.00 per million output tokens ($8.00 blended price under 3:1 weighting), matching Claude Opus 5.5 in token pricing while providing a value efficiency of $0.1575 per index point. Under rigorous multi-agent workloads, Muse Wonder streams tokens at a stable 88 tokens per second with an initial latency of 2.15 seconds.
Featuring native multimodal capabilities spanning code, architectural blueprints, UI design assets, and video walk-throughs alongside a 1,048,576-token context window, Muse Wonder is purpose-built for end-to-end full-stack application development, automated test scaffolding, and self-directed software debugging loops. Enterprise development teams rely on Muse Wonder for mission-critical software engineering pipelines that require dependable reasoning, automated unit test generation, and seamless cross-platform software integration.
