MENLO PARK, CALIFORNIA / RankWire.AI / – Meta has launched Muse Wonder and Muse Wonder Full, the most advanced artificial intelligence models it has developed so far. Designed specifically for coding, long-duration AI agents, and intricate personal or professional tasks, these systems aim to enhance user capabilities. Meta states that the models can assist with managing insurance claims, interpreting leases, analyzing financial data, and building functional applications. Muse Wonder was made available on October 12 via the Meta Model API and Muse Code, while Muse Wonder Full was released exclusively for Muse Code subscribers.
This release marks a significant expansion of Meta’s focus on AI systems capable of executing multi-step processes, moving beyond simple prompt responses. Muse Wonder is able to plan extended tasks, assign subtasks to sub-agents, integrate with software tools, and recover from errors during execution. The models are also optimized for advanced computer use across browsers, desktop programs, and mobile apps. Supporting a context window of roughly one million tokens, both models can handle extensive information within a single session.
Meta introduced two new features—Active Memory and Active Focus—to facilitate longer and more tailored task execution. Active Memory retains relevant details throughout conversations and retrieves information when necessary for new tasks. Active Focus enables the models to persist with difficult tasks, test alternative strategies, and recover from mistakes. Additionally, Meta notes that the models can request clarification from users and verify actions before proceeding with critical operations.
Models focused on coding and persistent AI agents
Muse Wonder is priced at $4 per million input tokens and $20 per million output tokens via the Meta Model API. Cached input costs are $0.20 per million tokens, as per the company’s developer resources. Muse Wonder Full is available only through paid Muse Code subscriptions, not separate API plans. Developers with higher workloads can opt for Power Usage at $50 a month or Maximum Usage at $200 per month.
Meta claims that Muse Wonder Full outperforms several benchmarks related to agentic behavior and coding. The company highlights tests like DeepSWE and SWE-Marathon, which evaluate software engineering and long-duration coding capabilities. An independent evaluator, Artificial Analysis, assigned Muse Wonder an Intelligence Index score of 52, and Muse Wonder Full scored 56 at maximum reasoning effort. These results place the models above Meta’s earlier Muse Spark generation on the evaluator’s index.
Enhanced safety features accompany increased model functionality
Meta has trained the Muse Wonder series to prevent sharing private information without consent and to decline harmful requests. The models also include cybersecurity features that assist developers in identifying vulnerabilities in their code and improving security measures. Meta states that the models are assessed under the Meta Superintelligence Scaling Framework and plans to publish a comprehensive report on additional safety evaluations following the initial launch.
About 14 months prior to the Muse Wonder debut, Meta undertook a major overhaul of its AI initiatives. The company emphasizes that the new models build upon the foundation established with Muse Spark, with training conducted at its Prometheus data center. Meta intends to integrate Muse Wonder into more experiences within its Muse platform, starting with subscribers. The October release provides immediate access for developers to Muse Wonder, reserving Muse Wonder Full for Muse Code subscribers.