Meta’s Muse Wonder Scores 46.1% on CursorBench at High Effort

Meta’s Muse Wonder is listed as a selectable model in Cursor and scored 46.1% on CursorBench 4.0 at the default High effort setting. Its Max setting scored 50.4%, while the highest listed result was 57.8%.

When checked at 3:39 a.m. ET (07:39 UTC) on Oct. 16, Muse Wonder’s High result showed 33,612 tokens over 84 steps and an average cost of $3.51 per task. The Max run used 49,692 tokens over 114 steps and averaged $5.25 per task.

Claude Opus 5.5 Max led the listed results. Its 57.8% score was 11.7 percentage points above Muse Wonder at High and 7.4 points above it at Max. Small score gaps may not represent meaningful performance differences because results can vary.

CursorBench 4.0 evaluates coding agents on tasks drawn from Cursor sessions. Introduced Sept. 10, the version uses longer tasks covering editing, refactoring, investigation, understanding intent, job management and adherence to design requirements. Its scores describe performance on that test set, not every kind of software work.

Muse Wonder is the successor to Muse Spark 1.3. It has a 300,000-token context window, expandable to 1 million tokens in Max Mode, and six effort levels, with High as the default. Pricing is $4 per million input tokens and $20 per million output tokens; usage is outside what Cursor plans include.

Meta introduced Muse Wonder and Muse Wonder Full on Oct. 12. Wonder is available through the Meta Model API and Muse Code, while Wonder Full is available to Muse Code subscribers. The benchmark result adds a new development to Meta’s Muse Wonder launch.