MENLO PARK, CALIFORNIA / RankWire.AI / – Meta has unveiled Muse Wonder and Muse Wonder Full, a new suite of sophisticated artificial intelligence models. These systems are designed for software development, AI agents, and tackling intricate daily challenges. They assist with tasks such as leases, insurance claims, financial data, and application creation. Meta launched Muse Wonder through its Model API and Muse Code on October 12. Muse Wonder Full became accessible to users with eligible Muse Code subscriptions.
The focus of these models is on multi-step tasks rather than simple, single-answer queries. Muse Wonder is capable of planning workflows, utilizing digital tools, and dividing large projects into smaller, specialized sub-agents. It can operate across various platforms including browsers, desktop software, and mobile apps. Meta designed the system to recover smoothly when a chosen method encounters failure. Supporting context windows of approximately one million tokens, the models can handle large documents, extensive codebases, and other substantial materials within a single session.
Meta has also introduced Active Memory and Active Focus alongside the Wonder models. Active Memory retains useful details from previous conversations and retrieves them when necessary. Active Focus enables an AI agent to persist through challenging problems and adapt its approach after encountering errors. The models can ask for clarification if instructions lack key information and can seek approval before executing sensitive or critical actions. These features aim to extend longer work sessions without the need for users to restart tasks repeatedly.
Expanded agent and coding functionalities with new models
Muse Wonder is priced at $4 per million input tokens via the Meta Model API. Output tokens cost $20 per million, while cached input tokens are billed at $0.20 per million. Instead of separate API pricing, Meta offers Muse Wonder Full through paid Muse Code plans. The company provides a Power Usage plan for $50 monthly, and a Maximum Usage plan at $200 per month, catering to developers with more intensive coding needs.
Meta has released benchmark data evaluating software engineering, coding agents, and long-term technical tasks. The results include tests like DeepSWE and SWE-Marathon. Additionally, Artificial Analysis assessed both models at maximum reasoning effort, awarding Muse Wonder an Intelligence Index score of 52 and Muse Wonder Full a score of 56. These scores surpass those of Meta’s earlier Muse Spark models, demonstrating an improved reasoning performance based on the evaluator’s scale.
Safety and privacy measures accompany the new models
Meta emphasizes that the Wonder models incorporate protections for privacy, harmful requests, and cybersecurity. The models are trained to avoid disclosing private information without user consent and to reject harmful activity requests, all while maintaining essential security functions. Developers are still able to utilize these models to identify vulnerabilities in their own software and enhance their defenses. Meta conducts safety evaluations of its most advanced AI systems under its Superintelligence Scaling Framework and plans to publish further details on safety testing for the new models.
This launch follows a significant restructuring of Meta’s AI efforts that began roughly 14 months earlier. The company states that Muse Wonder builds upon the technical foundation established by Muse Spark. The new models were trained using computing resources at Meta’s Prometheus data center. Additionally, Meta intends to integrate Muse Wonder technology into more features within the Muse platform, starting with subscribers. Developers can currently access Muse Wonder, while Muse Wonder Full remains available through qualifying Muse Code subscription plans.