Muse Wonder

VentureBeat reports that Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, 2026, targeting complex software engineering, legal, business automation, and long-running agent tasks. Neither model ships with downloadable weights; standard Wonder is offered commercially via API while Wonder Full is gated to Muse Code subscribers. Standard Wonder costs $4/$20 per million input/output tokens with cached input at $0.20 per million tokens. Artificial Analysis awarded Wonder Full a 56 on its Intelligence Index, tying Claude Sonnet 5.5 and surpassing GPT-6 Astra, Gemini 4 Argon, and Fable 5.1; standard Wonder earned 52, a material gain over Muse Spark 1.3's 48.

TechCrunch reports that Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on October 12, 2026, marking Meta's first frontier-model effort with larger, more capable successors to the lightweight Muse Spark. Per Meta's model card, both models achieve benchmark scores comparable to Fable 5.1 and GPT-6 Astra, emphasizing long-horizon active memory. Standard Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens via the Meta Model API, roughly on par with GPT-6 Sol or Opus 5.5. Muse Wonder Full is restricted to Muse Code subscribers and is not served through the API, while both models will also back the Muse agent.

Meta's corporate newsroom post confirms the launch of Muse Wonder and Muse Wonder Full, described as its most capable models to date with state-of-the-art performance on personal and professional agentic tasks. The standard Wonder is available to developers and builders through the Meta Model API and Muse Code, with Wonder Full exclusive to Muse Code subscribers. The series targets expert domains including law, finance, and health alongside frontier coding, and introduces Active Memory for retaining user context across conversations plus Active Focus for sustained task pursuit, error recovery, and clarification-seeking. Wonder will also power experiences in the Muse assistant, rolling out first to subscribers.

Meta AI Research introduced the Muse Wonder series, led by Muse Wonder and Muse Wonder Full, as the company's first frontier models for personal and professional agentic tasks. The standard Muse Wonder is available today through the Meta Model API and Muse Code, while Wonder Full is reserved for Muse Code subscribers, and the series will later power the Muse assistant and artifact generation. Trained at Meta's gigawatt-scale Prometheus data center, the series builds on the groundwork of Muse Spark and is designed around Active Memory and Active Focus capabilities for long-horizon tasks. Meta reports state-of-the-art results across benchmarks for personal and professional agents, memory, and coding, marking the next step on its scaling ladder toward personal superintelligence.

Meta announced Muse Wonder on October 12, 2026, as a frontier reasoning model developed by Meta Superintelligence Labs for complex coding, long-running agents, computer-use scenarios, and professional workflows. The base model accepts text, image, video, audio, and PDF documents and returns text, with a 1,048,576-token context window for long-horizon task execution. Artificial Analysis scored Muse Wonder at 52 on its Intelligence Index v4.3, compared with 48 for the prior Muse Spark 1.3, reflecting a meaningful jump in independent benchmark performance. The model is accessible through Meta's public API and the Muse Code product, with API pricing set at $4 per million input tokens and $20 per million output tokens.

llm-stats.com lists Muse Wonder with a blended price of $4.76 per million tokens alongside aggregated per-benchmark scores sourced from the Meta Muse Wonder model card dated October 12, 2026. It ranks first on MemoryBench at 0.86 and second on DeepSWE 1.1 at 0.79, among results drawn from Meta's official card. The aggregator also reports a third-place partial-credit OSWorld 2.1 score of 0.73 and positions Muse Wonder's LLM Stats Score at 47.6 against Claude Opus 5.5's 60.4. Benchmark sources are explicitly attributed to the Meta Muse Wonder model card rather than independent re-runs.