Meta ships Muse Wonder, an agentic model with a 1M-token context
Two new frontier models add persistent memory across conversations and the ability to keep working on a goal through setbacks, priced at $4 input and $20 output per million tokens.
Meta released Muse Wonder and Muse Wonder Full on 12 October, a pair of frontier models built for “personal and professional agents” rather than chat. Both take text, image, video, audio and PDF input and return text, with a 1,048,576-token context window.
The models add two named capabilities: Active Memory, which carries context about a user across separate conversations and retrieves it when relevant, and Active Focus, the model continuing to work toward a goal and trying alternative approaches when the first one fails. Meta says the models lead on long-running software projects and agentic coding benchmarks including SWE-Marathon. Muse Wonder is live now through Meta’s Model API at $4 per million input tokens and $20 per million output, with cached input at $0.20 per million. Muse Wonder Full is restricted to Muse Code subscribers on its Power Usage and Maximum Usage tiers.
Why it matters: persistent memory and goal persistence as named, documented features (rather than something you bolt on with your own scaffolding) are worth a look if you’re building anything that runs an agent over multiple sessions or needs it to recover from a failed attempt on its own. The 1M context window and per-token pricing put it in the same bracket as Claude and GPT-6 for cost comparisons.
The caveat: Meta’s own blog post doesn’t publish benchmark scores, only names of the evaluations used; the specific numbers so far come from third parties reading Meta’s API documentation, not from Meta’s announcement itself.