Meta's Muse Wonder AI model launch

Muse Wonder became available in Cursor on October 15, scoring 46.1% on CursorBench at default effort, the highest of any Muse model on that benchmark.

AI #190 flagged signs of benchmaxxing in Meta's presentation of the October 12 launch, noting Muse Spark 1.3 had the same independent rating as GPT-6 Sol despite likely being weaker in practice. Day 7 retention exceeded 30%, above ChatGPT's, but Day 14 dropped to around 22%, and AI #190 noted push notifications may be driving opens rather than genuine utility.

A video interview published October 14 named Meta's new AI initiative Meta Superintelligence Labs and added a $100B investment figure; Zuckerberg and Wang also discussed agentic commerce, what lies beyond the smartphone era, and why AI development will not produce a winner-take-all outcome.

Reporting on the October 12 release noted the development followed a significant capital reorientation beginning last August, and early reactions identified the pricing as a potential limit on developer adoption.

Meta released Muse Wonder and Muse Wonder Full on October 12, with Zuckerberg telling the Financial Times Meta has caught up with OpenAI and Anthropic at the AI frontier.

Meta's chart shows Muse Wonder Full winning 7 of 14 rows, but Artificial Analysis independently scores it at 56, below Claude Opus 5.5 at 58, and Wonder Full is not available via API. SensorTower shows Muse leading ChatGPT on Day 1 retention by 32 points but trailing by 4.3 at Day 14.

Simon Willison reported on October 14 that Muse Wonder, which Meta released on October 12, supports six reasoning levels from minimal to max, and that Artificial Analysis scores it at the max setting as equivalent to GPT-6.1 Sol, a higher mark than the 56 Intelligence Index figure cited for its default configuration.

Zuckerberg attributed Llama 4's underperformance to staffing it like a large product team rather than a small research group, describing the setback as an entrepreneurial test.