MENLO PARK, CALIFORNIA / RankWire.AI / – Meta has introduced Muse Wonder and Muse Wonder Full, its most capable artificial intelligence models to date. The company designed the new systems for coding, long-running AI agents and complex personal or professional tasks. Meta said the models can help users navigate insurance claims, understand leases, analyze finances and build working applications. Muse Wonder became available on October 12 through the Meta Model API and Muse Code, while Muse Wonder Full launched for Muse Code subscribers.
The release extends Meta’s push into AI systems that can complete multi-step work rather than only answer individual prompts. Muse Wonder can plan longer tasks, delegate work to sub-agents, use software tools and recover from errors during execution. Meta also built the models for advanced computer use across browsers, desktop programs and mobile applications. Both models support a context window of about one million tokens, allowing them to process large amounts of information in a single session.
Meta added two features called Active Memory and Active Focus to support longer and more personalized tasks. Active Memory carries relevant information across conversations and retrieves details when they matter to a new task. Active Focus helps the models continue working when a task becomes difficult, test another approach and recover after an error. Meta said the models can also ask users for clarification and check before completing important actions.
Models target coding and long-running AI agents
Muse Wonder costs $4 per million input tokens and $20 per million output tokens through the Meta Model API. Cached input costs $0.20 per million tokens, according to the company’s developer information. Meta offers Muse Wonder Full only through paid Muse Code subscription plans rather than separate API pricing. The company lists Power Usage at $50 a month and Maximum Usage at $200 a month for developers with heavier coding workloads.
Meta says Muse Wonder Full leads several agentic and coding benchmarks in results published with the release. The company highlights DeepSWE and SWE-Marathon among the tests used to measure software engineering and long-running coding performance. Independent evaluator Artificial Analysis gave Muse Wonder an Intelligence Index score of 52 and Muse Wonder Full a score of 56 at maximum reasoning effort. Those results place the models above Meta’s earlier Muse Spark generation on the evaluator’s index.
Safety controls accompany broader model capabilities
Meta trained the Muse Wonder family to avoid sharing private information without permission and to reject requests involving harmful acts. The company also preserved cybersecurity functions that can help developers identify flaws in their own code and strengthen defensive systems. Meta said it evaluates its most capable models under the Meta Superintelligence Scaling Framework. The company says it will publish a fuller report covering additional safety evaluations following the initial release.
The company began a broad overhaul of its artificial intelligence work about 14 months before the Muse Wonder launch. Meta says the new models build on foundations established with Muse Spark, with training conducted at its Prometheus data center. Meta plans to bring Muse Wonder into more experiences inside its Muse product, starting with subscribers. The October release gives developers immediate access to Muse Wonder while reserving the larger Muse Wonder Full model for Muse Code subscribers.