MENLO PARK, CALIFORNIA / CryptoWire / – Meta has released Muse Wonder and Muse Wonder Full, two advanced AI models built for coding and agent-based work. The models can handle complex tasks that unfold across several steps. Meta says they can assist with software development, insurance claims, lease reviews and financial information. Developers can access Muse Wonder through the Meta Model API and Muse Code. Meta offers Muse Wonder Full to users on eligible Muse Code subscription plans.
Meta built the Wonder models to complete longer assignments that require planning, tool use and repeated decisions. They can divide work among specialized sub-agents and interact with browsers, desktop programs and mobile applications. Muse Wonder can adjust its approach after an error instead of ending the task. Both models support context windows of about one million tokens. That allows them to work with large codebases, lengthy documents and other high-volume material in one session.
Meta also added Active Memory and Active Focus to the new model family. Active Memory keeps useful information from earlier conversations and recalls it when a later task requires that context. Active Focus helps the models continue through difficult assignments and try another method when progress stalls. The systems can request more information when instructions are incomplete. They can also seek user confirmation before carrying out important actions, adding another check during longer AI-driven workflows.
New AI models focus on developers and long tasks
Meta prices Muse Wonder at $4 per million input tokens through the Meta Model API. The company charges $20 per million output tokens and $0.20 per million cached input tokens. Meta distributes Muse Wonder Full through paid Muse Code plans rather than separate API access. Its Power Usage tier costs $50 a month. The Maximum Usage tier costs $200 a month and provides a higher usage allowance for developers working on more demanding coding tasks.
Meta published performance results for software engineering, reasoning and extended agent workflows. Its material includes benchmark tests such as DeepSWE and SWE-Marathon. Artificial Analysis separately evaluated both models at maximum reasoning effort. Muse Wonder received an Intelligence Index score of 52, while Muse Wonder Full scored 56. Those marks exceed the results recorded for Meta’s earlier Muse Spark generation on the same evaluation framework. The figures give developers another reference point for comparing the new models with Meta’s previous systems.
Safety framework covers privacy and cyber use
Meta says the Wonder models include protections for private data, harmful requests and cybersecurity use. The company trained them to avoid sharing private information without permission. The systems can refuse requests connected to harmful activity while preserving legitimate defensive security functions. Developers can use the models to examine their own software and look for weaknesses. Meta assesses advanced AI systems under its Superintelligence Scaling Framework and says it will release more detailed information on safety testing after the launch.
The release follows a broader reorganization of Meta’s artificial intelligence work that began about 14 months earlier. Meta says Muse Wonder builds on technical work developed through the Muse Spark family. The company trained the new models using computing infrastructure at its Prometheus data center. It also plans to bring Muse Wonder capabilities into more parts of the Muse product, starting with subscribers. Developers can already access Muse Wonder, while Meta limits Muse Wonder Full to qualifying Muse Code subscription plans.