Meta Muse Wonder: Frontier Models for Personal and Coding Agents

Meta just put a new agent stack on the API. On October 12, 2026, the company introduced Muse Wonder and Muse Wonder Full as its most capable models for personal and professional agentic work — with a matching developer launch note for Muse Code and the Meta Model API.

What shipped

• Muse Wonder on Muse Code and the Meta Model API at $4 input / $20 output per million tokens (USD), with cached input at $0.20
• Muse Wonder Full only through Muse Code Power Usage and a new Maximum Usage plan (80× the Everyday Usage limit)
• Focus areas Meta highlights: Active Memory, Active Focus, frontier coding, long-running agents and computer use
• 1M context window listed on the developer page; multimodal reasoning called out for screenshots, charts, documents, images and video

How Meta frames the models

The research post ties the release to Meta’s Muse Spark scaling path and training at the Prometheus gigawatt-scale data center. Active Memory is described as carrying important context across conversations; Active Focus as keeping work going through blockers while asking for clarification before important actions. Meta claims state-of-the-art results on personal/professional agent, memory and coding evals, and says Muse Wonder Full tops DeepSWE and SWE-Marathon in independent third-party runs cited in the developer post — treat those as vendor-reported unless you re-run them yourself.

Safety note

Meta says the Muse Wonder family leads on resistance to adversarial instructions and careful handling of private information, with stronger CBRN and cyber-misuse safeguards while keeping defensive cyber capabilities available under its Superintelligence Scaling Framework.

Rocks take

Another frontier lab shipping “agent” pricing next to a coding IDE subscription. Useful if you already live in Muse Code; for everyone else, the public API SKU is the part to watch against Claude and GPT agent rates.