Muse Wonder Full

Changes from previous version

Same-day subscription tier above Muse Wonder 1.0. No API token price. DeepSWE v1.1 pass@1 moves from 78.8% to 79.0%, and Terminal-Bench 4.0 accuracy from 52.0% to 57.1%. SWE-Marathon is reported at xhigh effort (57.5%) rather than the max effort used for 1.0 (53.1%).

Release Summary

Meta's highest Muse Wonder tier, included with Muse Code Power Usage and Maximum Usage. No per-token price. Text, image, video, and audio in, text out. Meta publishes a 1,048,576-token context and does not publish a separate output cap, so the catalog uses that context ceiling. At max effort, DeepSWE v1.1 pass@1 is 79.0% in Datacurve's runs and Terminal-Bench 4.0 accuracy is 57.1%. SWE-Marathon pass@1 is 57.5% at xhigh effort. API sibling, shipped the same day: Muse Wonder 1.0.

Timeline

Other Muse Wonder versions