Meta Muse Wonder Review: A Frontier Agent Model With Memory and Focus

On October 12, 2026, Meta Superintelligence Labs released two models, Muse Wonder and Muse Wonder Full, for personal and professional agent tasks. Meta markets two features: Active Memory (retaining context across sessions) and Active Focus (keeping multi-step agents on track as they switch between apps). The models were reportedly trained at Meta's Prometheus data center in Ohio. Unlike Llama, Muse Wonder is closed: it is available only through Meta's API, Muse Code and Meta apps, and Meta has not said whether weights will ever be released.

The following are published, verifiable figures with sources and test conditions. Independent third party Artificial Analysis Intelligence Index v4.3 (maximum reasoning effort, data as of October 13, 2026): Muse Wonder 52, Muse Wonder Full 56; Claude Opus 5.5 at 58, GPT-6 Astra at 53, Claude Sonnet 5.5 at 56 in the same period. Meta's published figures: 78.8 on DeepSWE 1.1 for Muse Wonder (Claude Opus 5.5 at 74.2), 79.0 for Muse Wonder Full; 53.1 on SWE-Marathon (Meta notes this is not a common-harness comparison); above Fable, Opus and Astra on MemoryBench (raw scores not published). Meta's internal composite Personal Superintelligence Index: Wonder 54.3, Wonder Full 59.1 (Opus 5.5 at 55.2, GPT-6 Astra at 54.5). Note: Artificial Analysis is independent testing; Meta's figures are vendor-published.

1. Pricing: Muse Wonder is billed at $4/$20 per million tokens (input/output) via the Meta Model API, $0.20 cached input; Meta has reportedly not officially confirmed all figures.
2. Variant split: Muse Wonder is on the API (muse-wonder-1.0) and Muse Code; Muse Wonder Full is for Muse Code subscribers only (Power $50/month, Maximum $200/month), not on the API.
3. Context: About 1 million tokens.
4. Build: About 14 months, led by Alexandr Wang's team.