MENLO PARK, CALIFORNIA / RankWire.AI / – Meta has launched Muse Wonder and Muse Wonder Full, its most powerful artificial intelligence models to date. Designed for programming, long-duration AI agents, and complex personal or professional tasks, these systems aim to assist users with insurance claims, lease comprehension, financial analysis, and application development, according to Meta. Muse Wonder was made accessible on October 12 via the Meta Model API and Muse Code, while Muse Wonder Full was released for Muse Code subscribers.
This launch signifies Meta’s expansion into AI systems capable of executing multi-step workflows, moving beyond simple prompt responses. Muse Wonder can plan extended tasks, assign subtasks to sub-agents, utilize software tools, and recover from errors during operation. The models are also optimized for advanced computing use across browsers, desktop software, and mobile apps. Both support a context window of approximately one million tokens, enabling the processing of substantial information within a single session.
Meta introduced two features—Active Memory and Active Focus—to facilitate longer, more tailored interactions. Active Memory maintains relevant details across conversations and retrieves information pertinent to new tasks, while Active Focus helps the models persist through challenging tasks, explore alternative methods, and recover from mistakes. Additionally, the models can request clarifications from users and verify actions before execution, Meta stated.
Focus on Coding and Persistent AI Agents
The pricing for Muse Wonder is $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 outlined in the company’s developer documentation. Muse Wonder Full is only available through paid Muse Code subscriptions, not separate API pricing. For developers with intensive coding needs, Meta offers Power Usage at $50 a month and Maximum Usage at $200 a month.
According to Meta, Muse Wonder Full outperforms many benchmarks related to agentic and coding tasks, with results published alongside the launch. Tests like DeepSWE and SWE-Marathon, which evaluate software engineering and long-duration coding, were used for assessment. Independent evaluator Artificial Analysis assigned Muse Wonder an Intelligence Index score of 52 and Muse Wonder Full a score of 56 at maximum reasoning effort. These scores surpass those of Meta’s earlier Muse Spark generation, as per the evaluator.
Enhanced Safety Features for Broader Capabilities
Meta has trained the Muse Wonder models to avoid sharing private information without user consent and to decline requests involving harmful activities. The models also retain cybersecurity features that help developers detect vulnerabilities in their code and improve defense mechanisms. Meta states it evaluates its most advanced models under the Meta Superintelligence Scaling Framework and plans to publish a comprehensive report on additional safety assessments following the initial release.
The company began overhauling its AI efforts approximately 14 months before the Muse Wonder debut, building on the foundations laid by Muse Spark, with training carried out at its Prometheus data center. Meta intends to integrate Muse Wonder into more features within its Muse platform, starting with subscribers. The October launch provides immediate access for developers to Muse Wonder, while the larger Muse Wonder Full remains exclusive to Muse Code subscribers.