Meta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It

Meta has shipped Muse Wonder and Muse Wonder Full, its newest models for long-running agent work. The pitch is broad: coding, legal and finance tasks, plus computer use. Muse Wonder is already live on the Meta Model API and in Muse Code, while Muse Wonder Full is fenced off behind paid Muse Code plans.

The headline feature is scale. Both models take a 1 million-token context and accept text, image, video and audio input, with text output only. Meta says the models run in its cloud, either through the API or through the Muse Code CLI on macOS, Linux and Windows. Wonder itself costs $4 per million input tokens and $20 per million output tokens, while Wonder Full comes only with Muse Code Power at $50 or Maximum at $200 a month.

Meta is leaning hard on agent behavior here. The company says Muse Wonder was trained at its Prometheus data center and improved through pretraining and reinforcement learning. It also points to a larger base model as the reason the system does better on unfamiliar tasks. Two named behaviors, Active Memory and Active Focus, are meant to keep context alive across conversations and keep difficult tasks moving, even if the model has to try alternatives and check in before doing something important.

On the benchmarks, the story is mixed rather than triumphant. Meta’s table says Muse Wonder Full leads 6 of 14 rows, but it still trails Claude Opus 5.5 in several places. Muse Wonder’s strongest score is 85.6 on MemoryBench, which Meta lists as the top result among the five models in the table. Its weakest showing is 50.0 on ProgramBench, far behind Claude Opus 5.5 at 70.5.

For builders, the practical path is straightforward. Sign in at dev.meta.ai, create an API key after adding billing, and test muse-wonder-1.0 in the Playground. Or install Muse Code, then switch models in the terminal with /models or launch directly with muse --model muse-wonder-1.0. Meta also says the API is OpenAI-compatible, so existing OpenAI SDK code can point at a new base URL and go to work.