What happened
On October 12, 2026, Meta introduced Muse Wonder and Muse Wonder Full, the first models in its Muse Wonder series and, by Meta’s description, its most capable models to date. They build on the earlier Muse Spark model and were trained at Meta’s Prometheus data center. Meta targets them at personal and professional agentic tasks: multi-step browser and cross-app workflows, coding, and domain work in law, finance and health.
What’s new
- Active Memory: the model keeps track of important context about the user across conversations and retrieves relevant details for the current task.
- Active Focus: on hard goals, the model tries alternative approaches when blocked, adapts to new information, asks for clarification when needed and checks in before important actions.
- Model ID and limits: muse-wonder-1.0 on the Meta Model API, with a 1,048,576-token context window. Inputs: text, image, video, audio and PDF. Output: text only. Reasoning effort is adjustable from “minimal” to “max”.
- Pricing (standard mode, per million tokens): $4.00 input, $0.20 cached input, $20.00 output.
- Rate limits: 3,000 requests and 4,000,000 tokens per minute, applied per team rather than per API key.
- Muse Wonder Full: available only to Muse Code subscribers and not served through the API.
Meta says the models reach state-of-the-art results on agent, memory and coding evaluations, naming Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon. The announcement text does not list the scores; they are in Meta’s chart and model card. Meta describes the models as having its strongest safeguards against CBRN and cyber misuse and says a full safety report is coming soon.
Why it matters
This is Meta’s entry into paid, closed frontier models served through its own developer API, rather than the open-weight Llama releases many developers associate with the company. At $4/$20 per million tokens it is priced in the same range as other flagship agentic models, and the 1M-token context plus native audio and video input make it a candidate for long-document and multimodal agent work.
Memory features also raise practical questions for builders: how persisted user context is stored, scoped and deleted. Teams handling personal data should review Meta’s API documentation on memory behavior before enabling it in production.
How to try it
- Create a project on the Meta Model API at dev.meta.ai and call muse-wonder-1.0.
- Use Muse Code (dev.meta.ai/code) to test coding agents; a subscription unlocks Muse Wonder Full.
- Meta says Muse Wonder will also power experiences in the Muse app, starting with subscribers.