Meta Superintelligence Labs' new agentic model lands in the popular AI code editor with a mid-pack benchmark score and a 1M-token context window
Meta Superintelligence Labs has a new model in one of the most competitive rooms in AI: the code editor. Muse Wonder is now available inside Cursor, and it posted a score of 46.1% on CursorBench.
That is a respectable debut, not a coronation. The current leader on the same benchmark sits more than 11 points higher, so Meta’s entry joins the race somewhere in the middle of the pack.
Cursor is the AI-powered code editor built by Anysphere. It lets developers choose which underlying model does the heavy lifting, which makes it a kind of open audition stage for AI labs.
Meta Superintelligence Labs released Muse Wonder on October 12, 2026, alongside a more capable sibling called Muse Wonder Full. Both are pitched at agentic work, meaning tasks where the model has to keep a goal in mind across many steps instead of answering one question and stopping.
To support that, Muse Wonder ships with a 1M-token context window. Tokens are the chunks of text a model reads, so a window that large lets it hold a sizable codebase or document set in working memory at once.
The model also accepts multimodal input, including text, images, video and audio.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Pricing through the Meta Model API is set at $4 per million input tokens and $20 per million output tokens .
The 46.1% figure comes from CursorBench 4.0, using the model’s ‘High’ configuration. Turned up to its ‘Max’ setting, Muse Wonder can reach as much as 50.4%.
For comparison, Claude Opus 5.5 leads at 57.8% in its maximum effort mode. Even at full throttle, Meta’s model trails that mark by 7.4 points.
The research summary describes the result as competitive among mid-tier models and favorable against several other frontier systems.
CursorBench itself is fairly new. Version 4.0 was introduced on September 10, 2026, roughly a month before Muse Wonder launched, and it focuses on long-horizon coding tasks drawn from real-world software development scenarios.
Meta’s own internal evaluations point to strengths in long-horizon coding and agent workflows.
The Muse Wonder models are built on Meta’s earlier Muse Spark series. They arrive as part of a broader effort to sharpen the company’s AI offerings following a significant restructuring of its AI division in mid-2025.