Meta Muse Wonder, a new family of AI models built for complex agentic work, debuted on Oct. 12, 2026, according to the company’s newsroom post. Meta says the series, which includes Muse Wonder and Muse Wonder Full, represents its most capable models to date and has closed the gap with frontier systems from OpenAI and Anthropic after more than a year and tens of billions of dollars of investment. The launch follows the April 8, 2026 unveiling of Muse Spark, the first model from Meta Superintelligence Labs, and signals a turn after the underwhelming April 5, 2025 debut of Llama 4. Both new models are proprietary, with no downloadable weights, and are now accessible to developers, builders, and subscribers through Meta’s various product lines.
Meta Muse Wonder Availability Across Meta Products
Meta says it is distributing the Muse Wonder family through several channels aimed at both developers and everyday users. Muse Wonder is available to developers and builders through the Meta Model API and Muse Code, while Muse Wonder Full is restricted to Muse Code subscribers. The standard Muse Wonder model will also power experiences inside Muse, Meta’s personal AI agent, and paying subscribers will gain access first before wider rollout. Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens through the API, while Muse Wonder Full is sold only through Muse Code at $50 a month for Power Usage and $200 a month for Maximum Usage. The company is pitching both models for personal and professional agentic tasks, listing examples such as handling an insurance claim, understanding a lease, making sense of finances, and building a working app. Meta says data handling differs by product: the API does not use submitted content for training, while Muse Code subscriptions default to a contributor mode that permits Meta to use prompts, code, and outputs to improve future models, with users able to opt out for future submissions.
Meta Muse Wonder Training Under Superintelligence Framework
Meta trained, evaluated, and released the Meta Muse Wonder models under its Meta Superintelligence Scaling Framework, with a fuller report containing additional safety evaluations expected in the coming period. The training took place at the company’s gigawatt-scale Prometheus data center, and the company says the models build on larger and stronger base models than Muse Spark 1.3, with Muse Wonder Full beating the Muse Spark 1.3 trend on unfamiliar tasks. Meta describes Wonder Full as its most aligned model to date, calling alignment central to how it builds personal superintelligence. The company says the models are trained to refuse requests that could help with harmful acts and to avoid sharing private information without permission, while Meta also built its strongest safeguards to date against chemical, biological, radiological, nuclear, and cyber misuse. Meta acknowledges that the models can still make mistakes on unfamiliar multistep tasks and notes it has not disclosed model sizes.
Meta Muse Wonder Benchmarks Against GPT-6 and Claude Opus
Meta Muse Wonder launches on the same date Meta confirmed its $145 billion 2026 AI capex commitment, marking the first major frontier release from Meta Superintelligence Labs since the April 2026 unveiling of Muse Spark. With Meta Muse Wonder tied to a closed-weight distribution model and priced below Claude Sonnet 5.5 on output tokens, the immediate question for enterprise buyers is not whether the model is competitive but how long Meta will keep Meta Muse Wonder proprietary before opening downloads to the open source community it once led. The Muse Wonder family ships on a 1 million token context window and accepts text, image, audio, and video inputs, putting Meta Muse Wonder in the same multimodal tier as GPT-6 Astra and Claude Fable 5.1 from launch day.
Meta Muse Wonder Strategy After Llama 4 and Scale AI Deal
Engadget reports that Meta says it has caught up with the frontier after more than a year of rebuilding and the company has said it may spend as much as $145 billion on AI this year. In an interview with the Financial Times, chief executive Mark Zuckerberg said the work required a large amount of compute and investment along with rigorous scientific work, and he described Wonder on X as a much larger and more advanced class of models than the earlier Muse Spark family. VentureBeat says Meta believes the models can outperform leading offerings from OpenAI and Anthropic in complex software engineering, legal work, business automation, and long-running agent tasks, with the company adding features such as Active Memory, which carries context across conversations, and Active Focus, which keeps the model working through hard goals by trying different approaches, working through blockers, asking for clarification, and checking in before important actions. The push follows the rocky reception of Llama 4, after which Meta released no new large language models for the following year, and a June 2025 deal in which Fortune reports Meta paid $14.3 billion for a 49 percent nonvoting stake in Scale AI while hiring founder Alexandr Wang as chief AI officer to lead Meta Superintelligence Labs, a move designed to put Meta Muse Wonder at the center of a more competitive AI operation.