MENLO PARK, CALIFORNIA / CryptoWire / – Meta has introduced Muse Wonder and Muse Wonder Full, its latest artificial intelligence models for coding and agent-based tasks. The systems can handle multi-step work across personal and professional settings. Meta says they can help users review leases, manage insurance-related tasks, examine financial information and build software. Muse Wonder launched through the Meta Model API and Muse Code. Muse Wonder Full is available through qualifying Muse Code subscription plans.
The new models can break larger assignments into smaller tasks and direct specialized sub-agents to complete parts of the work. They can also use browsers, desktop applications and mobile software during longer workflows. Muse Wonder can change methods when an approach fails and continue working after errors. Both models support context windows of about one million tokens. That capacity lets them process large documents, extensive code repositories and other lengthy material within a single session.
Meta equipped the Wonder family with features called Active Memory and Active Focus. Active Memory keeps relevant details from earlier conversations and retrieves them when needed during later tasks. Active Focus helps the models stay engaged with difficult work and try another path after setbacks. The systems can also ask users for missing information before continuing. They may request confirmation before important actions, giving users another control point during longer agent-driven workflows.
Wonder models broaden coding and agent features
Muse Wonder costs $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 offers Muse Wonder Full through paid Muse Code tiers rather than separate API pricing. Its Power Usage plan costs $50 per month, while the Maximum Usage plan costs $200 per month. Those plans provide different usage levels for developers working with the coding service.
Meta released benchmark results that cover software development, reasoning and long-running agent tasks. The company cited tests including DeepSWE and SWE-Marathon in its performance material. Artificial Analysis also tested the models at maximum reasoning effort. The evaluator assigned Muse Wonder an Intelligence Index score of 52. Muse Wonder Full received a score of 56. Both results stand above Meta’s earlier Muse Spark generation on the same independent measurement system.
Meta outlines privacy and safety protections
The company says Muse Wonder includes safeguards for privacy, cybersecurity and harmful requests. Meta trained the models to avoid releasing private information without permission. The systems can reject requests tied to harmful activities while retaining legitimate defensive security functions. Developers can use them to review their own code and identify software weaknesses. Meta assesses its advanced models under the Meta Superintelligence Scaling Framework and says more detailed safety evaluation information will follow the initial release.
The Wonder launch follows a wider reorganization of Meta’s artificial intelligence operations that began about 14 months earlier. Meta says the models build on work developed through its previous Muse Spark systems. The company trained the new generation with computing resources at its Prometheus data center. Meta also plans to bring Muse Wonder capabilities into more parts of the Muse product, beginning with subscribers. Developers can already access Muse Wonder, while Muse Wonder Full remains tied to eligible Muse Code subscriptions.