Meta launches Muse Wonder models for long-running personal and professional agents, giving developers a new model through the Meta Model API and a more capable subscription-only version inside Muse Code. The October 12 release adds persistent memory, sustained task execution and computer-use capabilities to Meta’s rebuilt AI portfolio.

 

Muse Wonder costs $4 per million input tokens and $20 per million output tokens, while cached input costs $0.20 per million tokens. Muse Wonder Full is not available through the public API at launch; access is limited to Muse Code’s Power Usage and Maximum Usage subscription plans.

 

The launch establishes five practical boundaries:

  • Two models with different access routes
  • One-million-token context windows
  • Memory designed to span conversations
  • Tools for coding and computer use
  • A full safety report still pending

 

Further Reading

 

Meta Launches Muse Wonder Models Across API and Muse Code

Muse Wonder is available as model identifier muse-wonder-1.0 through Meta’s OpenAI-compatible Model API and inside Muse Code. Meta says developers can switch existing OpenAI SDK projects by changing the base URL, retaining their client code while directing requests to its service.

 

The standard model carries a one-million-token context window and supports multimodal reasoning over screenshots, charts, documents, images and video. Meta says audio-dependent applications should confirm audio availability on the specific release surface rather than assuming every modality works through every product.

 

Muse Wonder Full is positioned above the API model for complex coding and agent workflows. It is included in Muse Code’s Power Usage plan and the Maximum Usage plan, which Meta says provides 80 times the Everyday tier’s limit. The company has not announced API pricing or a public API date for Full.

 

The models will also move into Meta’s Muse personal agent, initially for subscribers. That expansion connects the release to a consumer product that can work across browsers and applications, rather than leaving Wonder as a developer-only model family.

 

Active Memory and Active Focus Target Long-Running Work

Meta describes Active Memory as a system for carrying important context across conversations and retrieving it when a later task makes the information relevant. The design is intended to preserve goals, preferences, constraints and earlier decisions without requiring users to repeat the same background in every session.

 

Active Focus addresses a different failure mode: agents abandoning difficult work or repeating an unsuccessful approach. Meta says Wonder is trained to try alternatives, recover from tool errors, adapt to new information and request clarification when needed. It is also trained to check before taking important actions.

 

Those features support multi-file software changes, debugging, sequential tool calls and tasks that run through repeated testing cycles. They also raise a demanding product question. A useful agent must distinguish durable preferences from temporary instructions and avoid carrying sensitive or outdated information into an unrelated job.

 

Meta’s demonstrations include planning across browser applications, building a game from image inputs and creating a seating chart under multiple constraints. These examples illustrate intended capabilities, not measured production reliability. Real deployments will need audit logs, memory controls and clear recovery paths when an agent misreads context.

 

Meta’s Benchmark Claims Need Independent Replication

Meta presents Muse Wonder Full as its most capable model and reports strong results across professional work, coding, memory and computer-use evaluations. The published suite includes Harvey’s legal-agent benchmark, OSWorld, MemoryBench, Terminal-Bench, DeepSWE and SWE-Marathon, alongside Meta’s internal Personal Superintelligence Index.

 

The developer release says Full leads DeepSWE and SWE-Marathon in independent third-party runs. It also reports that the smaller Muse Wonder delivers performance close to Full on several tests, which would make the API model more commercially important if developers reproduce those results at its listed price.

 

Benchmark tables do not settle how the models behave inside long-lived agents. Scores can depend on prompts, reasoning budgets, tool scaffolds and grading rules. Meta’s internal composite is particularly difficult to compare outside the company, while the third-party coding results offer a stronger but still workload-specific signal.

 

The next evidence should come from independent tests using the released API model. Completion rates, human intervention, tool failures, latency and total cost over an entire task will reveal more than a single benchmark score about whether Wonder can sustain useful work.

 

Muse Wonder’s Safety Report Remains Unpublished

Meta says the Wonder family was trained to handle private information carefully, resist adversarial instructions and block assistance for harmful acts. It also says the model keeps advanced cybersecurity capabilities available for defensive work while applying stronger safeguards against chemical, biological, radiological, nuclear and cyber misuse.

 

The company has not yet published the full additional safety evaluation promised alongside the launch. That leaves important evidence unavailable at release, including detailed failure rates for privacy handling, prompt injection, dangerous capability thresholds and actions involving connected messages, calendars and documents.

 

Persistent memory increases the stakes because a model can accumulate context across time. Developers will need retention settings, deletion mechanisms, permission boundaries and records showing which stored detail influenced an action. Model-level training cannot replace application controls around accounts, tools and sensitive data.

 

Muse Wonder gives Meta a current, priced API model and a higher-end agent model inside its coding product. Its commercial impact will depend on whether Active Memory and Active Focus improve task completion without creating unmanageable privacy or oversight costs, and whether Meta’s pending safety report supplies evidence equal to the models’ broader access.