Meta Releases Muse Wonder With Active Memory And Focus
The API-tier model Meta shipped scores below GPT-6 Astra and Claude Opus 5.5 on its own composite index, so the practical-work framing and the memory and coding wins carry the launch rather than an overall lead.
Reporting from 1 source: GameBusiness.jp.
Meta announced the Muse Wonder and Muse Wonder Full AI models on October 12 US time, both already available through the Meta Model API and the Muse Code coding tool. Meta says Wonder scored 85.6 on MemoryBench and, paired with Muse Code, 68 on Artificial Analysis's coding agent evaluation. Chief AI Officer Alexandr Wang said the company aimed for practical work, not 3D demos, and the model card lists 3D spatial reasoning as a known limitation.
Both models are already live. Wonder Full is sold on Muse Code plans at $50 per month or more, and Meta says it reaches subscribers of the personal AI agent app Muse soon. The two features doing the work are Active Memory, which decides on its own when past information matters, and Active Focus, which tries other approaches when the model gets stuck.
Meta's internal MemoryBench score of 85.6 sits above Claude Opus 5.5's 81.5 and GPT-6 Astra's 81.4. The picture narrows elsewhere: Wonder scored 54.3 on Meta's Personal Superintelligence Index against Astra's 54.5 and Opus 5.5's 55.2, and Meta estimates $4.52 per task for Wonder.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.