Meta Muse Wonder AI: Capabilities and Use Cases for Law and Coding

How do Meta agentic AI models handle complex tasks?

Muse Wonder and Muse Wonder Full are two new AI models Meta launched on October 12, 2026, designed to handle complex tasks in professional fields including law, finance, health, and coding. These models represent a step toward personal superintelligence and are built upon the foundation of the earlier Muse Spark model.

The models were developed at Meta Superintelligence Labs and trained at the Prometheus data center in central Ohio. This work took place under an AI division led by chief AI officer Alexandr Wang, who was hired in June 2025 and restructured Meta’s AI efforts in August 2025.

Muse Wonder is the faster, more widely available version accessible through Meta’s API and Muse Code. Muse Wonder Full is the stronger model, built specifically for agentic work where the AI can plan, use tools, check its own work, and fix mistakes. While the standard Muse Wonder is broadly accessible, Muse Wonder Full is available only to Muse Code subscribers.

How do these models handle complex agentic tasks?

These models manage multi-step workflows like reviewing leases, handling insurance claims, and building apps through specific agentic features. They include "Active Memory," which maintains context across different conversations.

Additionally, "Active Focus" helps the models stay on task by working through errors and verifying actions before execution. You can process text, images, video, audio, and PDF inputs within a 1 million-token context window and receive text outputs. These models were developed under the internal codename "Watermelon," a detail teased by Mark Zuckerberg during an Instagram livestream before the launch.

Muse Wonder vs Muse Wonder Full: Which model should you use?

Performance metrics from the Artificial Analysis Intelligence Index show Muse Wonder Full scoring 56 and the standard Muse Wonder scoring 52, both outperforming Meta's earlier Muse Spark 1.3 model, which scored 48. Muse Wonder Full achieved a top-three placement on six of the ten tests in this index and uses 10% fewer output tokens per task than the older model.

According to the Artificial Analysis index reported by VentureBeat, Muse Wonder Full scores just below Claude Opus 5.5 but above Gemini 4 Argon, while the standard Muse Wonder scores below Gemini 4 Argon. In specific legal agent testing, Meta reported that Muse Wonder Full achieved a 13.1% full pass rate, compared to 4.2% for Claude Opus 5.5. Meta's model card compares these models against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5, asserting state-of-the-art results for agentic work in legal and long coding tasks.

How much do the Muse Wonder models cost?

The Muse Wonder API costs $4 per million input tokens and $20 per million output tokens. You also pay $0.20 per million cached input tokens, and Meta applies no surcharge for long context.

If you require web search functionality through the Meta Model API, the service costs $2.50 per 1,000 searches.

What are the core Muse Wonder Full capabilities?

You can access Muse Wonder through the Meta Model API as muse-wonder-1.0 in preview, or use it within Meta's coding agent, Muse Code. Muse Wonder Full is not available on the API and is restricted to the two highest usage plans of the Muse Code tool.

Meta also plans to integrate Muse Wonder into its personal assistant app, Muse, for subscribers in the near future.

What are the limitations and safety considerations?

Neither Muse Wonder nor Muse Wonder Full provides open weights, as Meta keeps these models closed and offers a separate open-weight model called Muse Glimmer.

The company has announced that a full safety evaluation report for these new models will follow.