Meta's Muse Wonder Arrives in Two Tiers, Then Trains the Models Most Users Actually Get

Meta launched its Muse Wonder frontier model series on October 12, 2026, releasing two variants: Muse Wonder, available via API and all Muse Code plans, and Muse Wonder Full, locked to the two most expensive Muse Code subscriptions. Three weeks later, on November 6, the company said it had used Wonder to train new "specialized" versions of the models already powering its consumer-facing Muse and Meta AI apps — adding the series' signature Active Memory and Active Focus features — while keeping the flagship Wonder models themselves out of reach for most users.

Two models, one API

Muse Wonder and Muse Wonder Full share a 1-million-token context window and multimodal input (text, image, video, audio; text output), but diverge on access and price [1]. Muse Wonder (muse-wonder-1.0) is sold on the Meta Model API at $4 per million input tokens, $0.20 per million cached input tokens, and $20 per million output tokens — the same list price as Claude Opus 5.5 [1]. It is also included in every Muse Code tier, from the $5/month Everyday Usage plan up to the $200/month Maximum Usage plan [1].

Muse Wonder Full (muse-wonder-full-1.0) has no per-token price. It is available only in Muse Code's Power Usage ($50/month) and Maximum Usage ($200/month) plans [1][9]. Meta's developer blog calls Wonder Full "our most advanced model yet" and says it "tops coding benchmark leaderboards, including DeepSWE and SWE-Marathon in independent third-party runs" [9]. As of November 3, developers could not build products on Wonder Full; the API serves only the smaller Wonder [1].

Both models are proprietary. Meta's open-weight offering remains Muse Spark Open, released October 15 under a commercial license with attribution requirements [1][5].

Benchmarks: Meta's table vs. independent indexes

Meta's own benchmark table, published at launch, shows Wonder Full leading rivals on several agentic and coding tests: Harvey LAB-AA (legal agent work), DeepSWE 1.1 and SWE-Marathon (long-horizon coding), MemoryBench (context retention), and Vals Finance Agent V2 [1]. On ProgramBench (writing programs from scratch), however, Claude Opus 5.5 leads at 70.5% against Wonder Full's 59.0% [1]. Opus 5.5 also wins HealthBench Pro, JobBench, Terminal-Bench, ProgramBench, and OSWorld's partial score in Meta's table [1].

Independent evaluations tell a different story. On Artificial Analysis's Intelligence Index v5 (updated October 27), Wonder Full scores 50 and Wonder scores 43 — behind Gemini 4 Carbon (59), Claude Opus 5.5 (58), and GPT-6 Astra (57) [1]. The Respan Index ranks Wonder Full fifth overall at 70.9 (coding 62.9, agents 74.1) but notes its reasoning and knowledge scores are estimated; the model has no public API [5]. BenchLM's live leaderboard (November 6) places Gemini 4 Carbon, Claude Opus 5.5, and GPT-6 Astra at the top; Muse Wonder does not appear in the top 10 [6].

Meta's launch table used an older version of the Artificial Analysis index; the v5 refresh makes direct comparison impossible [1]. The company also notes that Wonder Full's SWE-Marathon score used an "xhigh" reasoning setting, one step below the maximum used for other rows, and "isn't a same-harness comparison" [1]. Independent indexes also update on different schedules, so version dates matter across all comparisons.

Active Memory and Active Focus: claimed, not independently verified

Meta built two features into both Wonder models. Active Memory is designed to "carry what matters to you across conversations and bring back the details that are relevant to the task at hand," distinguishing temporary from permanent information and reasoning across time [8][1]. Active Focus aims to keep agents working through blockers, trying alternative approaches, and checking in before important actions [8][1].

Meta's MemoryBench shows both Wonder models at ~85.5%, ahead of Opus 5.5 (81.5%) and GPT-6 Astra (81.4%) [1]. The company has not published technical details of how either feature works, and independent tests are limited [1]. Audio support, listed as a capability, carries a caveat: "audio-dependent use cases require confirmation of audio support on the release surface" [1][9].

November 6: Wonder trains the mass-market models

On November 6, Meta announced it had trained new "specialized versions of our existing models" using Muse Wonder and was rolling them out to Muse and Meta AI users [10][3]. This is not a rollout of Muse Wonder itself. The company says the new models inherit Active Memory and Active Focus, tuned for consumer use cases: remembering which past details still matter, persisting through task setbacks, accepting mid-task clarifications, and deciding when to work independently versus collaboratively [10][7].

For Muse, Meta reports a 2.4-fold reduction in forgetting critical information, an increase in tasks correctly using critical context from 76% to 90%, a Personal Superintelligence Index rise from 38% to 47%, and a 67% improvement in completing everyday tasks with unexpected problems [10][7]. These are company-reported Muse results; no independent evaluation is provided, and Meta gives no separate numerical results for Meta AI [10][7]. The underlying completion rates for the 67% improvement figure are not disclosed [7].

What remains unknown

• Safety reporting: Meta's October newsroom post said a full safety report "will be available soon" [8]; as of November 6, the research blog still references the upcoming report without linking it [1].
• Muse assistant integration: Wonder "will soon power more of Muse, rolling out first to subscribers" [8][1]; no date has been given.
• Audio availability: Support must be confirmed per surface [1][9].
• Training method for specialized models: Meta "hasn't detailed its method" for distilling Wonder into the consumer models [3].

Assessment

Muse Wonder Full demonstrates genuine strength on long-horizon coding and legal agent benchmarks, both in Meta's table and in some independent runs (DeepSWE, SWE-Marathon). Muse Wonder offers that capability tier at a competitive API price ($4/$20) with notable speed (~196 output tokens/second on Artificial Analysis) [1]. However, the flagship model remains gated behind a $50–$200/month coding-agent subscription, and the mass-market rollout transfers only two labeled features — Active Memory and Active Focus — into models whose independent performance is unmeasured. The benchmark gap between Meta's launch figures and current independent indexes underscores the need to check which evaluation version underlies any claimed lead.