Muse Wonder: Models, Pricing, Benchmarks and Access

What is Muse Wonder?

Muse Wonder is a family of frontier artificial intelligence models developed by Meta's Superintelligence Lab. The lineup includes two models: the standard Muse Wonder and the larger, more capable Muse Wonder Full.

Meta trained both models at its gigawatt-scale Prometheus data center in central Ohio. While the previous Muse Spark focused on lightweight, low-latency execution, the Muse Wonder models target complex agent tasks and coding.

Key features of the Muse Wonder series

Muse Wonder introduces Active Memory and Active Focus to support long-running agent workflows. These systems allow the model to retain user context and work persistently toward goals.

• Active Memory: Retains user preferences and critical context across different conversations over days, weeks, and months.
• Active Focus: A looping system that tests alternative methods when an initial attempt fails, asks for clarification when necessary, and checks before performing sensitive actions.
• Domain expertise: Evaluated on specialized legal, financial, and medical tasks, including Harvey's Legal Agent Benchmark.
• Frontier coding: Handles multi-file codebases, iterative testing, and complex builds, such as generating playable 3D games.

Meta also showed Muse Wonder executing practical consumer workflows directly in browsers and apps. These demonstrations included building personalized running plans, arranging wedding seating charts, and ordering merchandise from social media photos.

Muse Wonder benchmark performance

Muse Wonder achieved high marks across industry evaluations for memory, automation, and software engineering. According to Meta, the models match or exceed competing offerings from Anthropic, OpenAI, and Google.

On the Artificial Intelligence Index, Muse Wonder Full placed ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon, trailing only Claude Opus 5.5. Meta reported that the model ranked first on the Personal SuperIntelligence Index, Automation Bench, and MemoryBench.

In coding evaluations, both Muse Wonder models outperformed Claude Opus and GPT-6 at their highest performance tiers. Meta also evaluates sustained software execution using the SWE-Marathon benchmark.

How much does Muse Wonder cost?

Muse Wonder is available through the Meta Model API with usage-based token pricing. Meta sets the pricing at roughly the same level as GPT-6 Sol and Claude Opus 5.5.

Muse Wonder Full does not have an API pricing tier. Meta does not offer Muse Wonder Full through the public API, reserving it instead for Muse Code subscribers.

How to access and use Muse Wonder

Muse Wonder access depends on whether you are a developer using an API or an end user working in Meta's applications. Meta provides different distribution channels for each variant in the series.

Developers can integrate Muse Wonder immediately by calling the Meta Model API. Muse Code subscribers can access both Muse Wonder and Muse Wonder Full directly inside the tool, while standard Muse consumer experiences will receive access in a staged rollout.

Safety protections and model limits

Muse Wonder includes alignment protections designed to handle private user data and prevent harmful actions. Meta built safeguards to protect against unauthorized data sharing and dangerous misuse.

The models are trained to refuse assistance with harmful requests, protect against cyber attacks, and resist adversarial prompts. Meta also implemented safeguards against chemical, biological, radiological, and nuclear threats under its scaling framework.

Frequently asked questions

Sources: Digital Trends · TechCrunch · research.meta.ai