Meta: Muse Wonder

Tech Jack Solutions' October 12, 2026 brief confirms Meta's first frontier-class model release and the price positioning: Wonder at $4.00 per million input tokens and $20.00 per million output tokens via the Meta Model API, with Wonder Full exclusive to Muse Code subscribers. Both were trained at Meta's Prometheus data center in central Ohio and power the Muse agent, succeeding the lighter Muse Spark. The piece explicitly flags that all cited benchmarks and MemoryBench results are Meta self-reported with no independent evaluation referenced. MemoryBench is highlighted as the headline differentiator for long-horizon retention, with Meta claiming scores higher than Fable 5.1, Opus 5.5, and GPT-6 Astra. Parameter counts are not disclosed, and the structured fact-box treats the entire benchmark comparison as a vendor claim rather than verified result. The brief positions Wonder Full as the consumer-only tier unavailable through any API endpoint.

Firstpost's October 12, 2026 coverage confirms the Muse Wonder and Muse Wonder Full launch through Meta Superintelligence Labs and elaborates on Active Memory's role in retaining relevant context across conversations. Meta positions them at personal and professional agentic tasks including lease comprehension, financial questions, and health information. The article notes Meta claims the models outperform GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 on select specialized benchmarks in legal document analysis and software development. Availability is split: Muse Wonder is accessible now via the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers. The coverage cautions that benchmark results may not always reflect real-world performance and that Meta's claims have not yet been independently verified. Evaluations reportedly covered domain expertise, Active Memory, Active Focus, and frontier coding capabilities.

Llm-stats consolidates Meta's self-reported scores from the October 12, 2026 Muse Wonder model card at max reasoning effort: MemoryBench 0.86 (rank 1), DeepSWE 1.1 at 0.79, and OSWorld 2.1 partial at 0.73. Additional rankings cover SWE-Marathon, JobBench, HealthBench, Vals Finance Agent v2, Terminal-Bench 4.0-AA, Program Bench, FocusBench, AutomationBench-AA, and Harvey LAB-AA. All methodology notes attribute the source to the Meta Muse Wonder model card rather than independent runs. Cost-efficiency scoring places Muse Wonder at an LLM Stats Score of 47.6 against Opus 5.5's 60.4, both at an identical $4.76 blended price per million tokens. A performance-by-conversation-depth breakdown tracks how Muse Wonder holds up as context extends. The page is gated behind a human-verification CAPTCHA, so only the excerpt is verifiable, and no third-party re-runs are cited.

An independent row-by-row audit of Meta's own benchmark table notes that only Muse Wonder is API-accessible, served as muse-wonder-1.0 on the Meta Model API, while Muse Wonder Full is reserved for Muse Code subscribers and is not API-served. At matched $4 in and $20 out pricing, the API-available Wonder loses to Claude Opus 5.5 on 9 of 14 of Meta's own benchmark rows. The model card lists a 1M-token context accepting text, image, video, and audio input. The analysis highlights that cached input is reported at $0.20 per million tokens, identical to Opus 5.5, while GPT-6 Astra and Claude Fable 5.1 list higher at $10 in and $50 out. Every cited number traces to Meta's model card, Artificial Analysis, or Meta's pricing source, with no independent evaluation. The piece underscores the gap between Meta's headline-grabbing Full variant and the API-available base Wonder for routing decisions.

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, 2026, calling them its first frontier-class models trained at the Prometheus gigawatt-scale data center. The series centers on Active Memory, which carries context across conversations, and Active Focus, which helps agents sustain long tasks and consult users before important actions. Meta cites top MemoryBench scores and claims out-of-distribution generalization improvements over Muse Spark 1.3. Muse Wonder is served via the Meta Model API and Muse Code at $4 per million input tokens and $20 per million output tokens, matching Claude Opus 5.5 pricing. Meta's own scorecard reports Muse Wonder Full ahead of Opus 5.5 on 8 of 14 benchmarks, with a SWE-Marathon gap of 57.5 to 46.3 and DeepSWE 1.1 at 79.0 versus 74.2. All benchmark numbers originate from Meta's self-reported model card with no independent third-party verification cited among the candidates.