Meta Muse Wonder Launches Frontier AI Models for Long-Horizon Agent Tasks

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026, via Meta Superintelligence Labs. The models target long-horizon agentic workflows such as insurance claims navigation, lease review, financial analysis, and app or game development.

Core Features of Meta Muse Wonder

The new models accept multimodal inputs including text, images, video, audio, and PDFs. They maintain an approximately one-million-token context window and generate text outputs. Two distinctive mechanisms stand out: Active Memory retrieves relevant user context across separate sessions, while Active Focus maintains goal-directed progress, tests alternative approaches, recovers from errors, requests clarification when needed, and seeks confirmation before executing major actions.

These capabilities differentiate the models from earlier Muse Spark releases that debuted in April 2026. Training occurred at Meta’s gigawatt-scale Prometheus data center as part of a 14-month internal AI overhaul. The weights remain closed, consistent with Meta’s updated Superintelligence Scaling Framework.

Practical use cases include sustained multi-step professional tasks where agents must track evolving requirements over hours or days. Enterprises can test these behaviors through the Meta Model API or Muse Code environments before broader deployment.

Benchmarks, Pricing, and Availability

Muse Wonder Full recorded a score of 56 on the Artificial Analysis Intelligence Index v4.3, placing it second behind Claude Opus 5.5. The standard Muse Wonder variant scored 52. Both models performed strongly on MemoryBench, DeepSWE, and SWE-Marathon long-horizon coding benchmarks, while results on OSWorld and ProgramBench remained mixed. Meta’s internal Personal Superintelligence Index gave Muse Wonder Full a 59.1 rating.

Pricing on the Meta Model API stands at $4 per million input tokens and $20 per million output tokens, with cached input available at $0.20 per million. Muse Wonder (model ID muse-wonder-1.0 preview) is accessible immediately. Muse Wonder Full requires a Power or Maximum subscription to Muse Code.

Safety evaluations continue, with a comprehensive report scheduled for release in the coming weeks. The models follow Meta’s established internal review processes but have not yet received external red-teaming at the scale of prior open releases.

Integration Plans and Next Steps

Muse Wonder will soon drive experiences inside Meta’s personal agent product named Muse, with initial access limited to subscribers. The same models will also support artifact generation and higher-intelligence workflows within the Muse platform. Specialized derivatives are under development to serve a wider range of Muse users over time.

No additional model release dates have been announced. Meta continues scaling efforts toward personal superintelligence while monitoring performance on agentic benchmarks. Developers can begin evaluation today through the API or Muse Code to assess fit for specific long-horizon workflows.