MENLO PARK, CALIFORNIA / RankWire.AI / – Meta unveiled its newest family of advanced artificial intelligence models, Muse Wonder and Muse Wonder Full, on October 12. Developed for tasks involving software development, AI agents, and complex daily activities, these models can assist in areas such as leases, insurance claims, financial data, and application creation. The company launched Muse Wonder via its Model API and Muse Code, with Muse Wonder Full becoming accessible to users holding eligible Muse Code subscriptions.
Focusing on multi-step tasks rather than single responses, the models are designed to plan workflows, utilize digital tools, and break down large projects among specialized sub-agents. They operate seamlessly across browsers, desktop applications, and mobile platforms. Meta engineered the system to recover gracefully if a selected approach fails. Supporting context windows of approximately one million tokens, the models can process extensive documents, codebases, and other large materials during a single session.
Additionally, Meta introduced Active Memory and Active Focus alongside the Wonder models. Active Memory retains relevant details from earlier conversations and can retrieve them when needed. Active Focus aids AI agents in tackling difficult problems and adjusting their strategies after encountering errors. The models are capable of requesting clarification when instructions lack critical details and can seek confirmation before executing sensitive or significant actions. These features aim to enable longer work sessions without users needing to restart tasks repeatedly.
Enhanced agent and coding features mark the new models
Pricing for Muse Wonder is set at $4 per million input tokens via the Meta Model API, with output costing $20 per million tokens. Cached input is billed at $0.20 per million tokens. Instead of separate API pricing, Meta offers Muse Wonder Full through paid Muse Code plans. The company also provides a Power Usage plan at $50 per month and a Maximum Usage plan at $200 per month, designed for developers requiring increased access for intensive coding tasks.
Meta shared benchmark results evaluating software engineering, coding agents, and long-term technical projects. The tests included DeepSWE and SWE-Marathon, with the models also assessed at maximum reasoning effort. Muse Wonder earned an Intelligence Index score of 52, while Muse Wonder Full scored 56. These results surpass Meta’s earlier Muse Spark models on the evaluator’s scale, indicating a clear improvement in reasoning capabilities.
Safety and privacy measures accompany the rollout
Meta states that the Wonder models incorporate protections for privacy, against harmful requests, and for cybersecurity purposes. They are trained to avoid revealing private information without user approval and to decline requests involving harmful activities, while still supporting legitimate security functions. Developers can utilize the models to identify vulnerabilities in their own systems and enhance security measures. Meta is evaluating its most advanced AI models under its Superintelligence Scaling Framework and plans to publish additional details regarding safety testing for these new systems.
This release follows a major restructuring of Meta’s AI efforts that began approximately 14 months earlier. The company explains that Muse Wonder builds upon the technical foundations laid by Muse Spark. Development was carried out using computing resources at its Prometheus data center. Moving forward, Meta intends to integrate Muse Wonder technology into additional features within the Muse platform, initially for subscribers. While developers can already access Muse Wonder, Muse Wonder Full remains available through eligible Muse Code subscription plans.