Meta's New Muse Wonder AI Model Can Remember and Stay Focused
Meta introduced two new AI models, Muse Wonder and a larger Muse Wonder Full, on October 12, 2026, built around persistent context rather than a single new benchmark claim. The models are centered on two features Meta calls Active Memory and Active Focus, aimed at document work, coding, and tasks in finance, law, and health, the kind of work where a model needs to track details across many steps rather than answer one question and stop. Meta says both were trained at Prometheus, its gigawatt-scale data center built for frontier AI work.
⚡ Quick facts
- What's new: Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026, built around two features called Active Memory and Active Focus
- Active Memory: carries relevant context forward across a conversation instead of losing it each session
- Active Focus: keeps the model working through multi-step tasks, recovers from its own errors, and checks in before consequential actions
- Access and pricing: Meta Model API at $4 per million input tokens and $20 per million output tokens, plus Muse Code; Wonder Full needs a Muse Code subscription
- Not yet confirmed: consumer Muse app subscription pricing for Wonder Full, and a rollout timeline beyond Meta's own API and Muse Code
What Meta announced
Meta introduced Muse Wonder and the larger Muse Wonder Full on October 12, 2026, in a post on Meta's research blog. Rather than leading with a single benchmark number, Meta framed the release around how well the models handle long, multi-step work: document tasks, coding, and work in fields like finance, law, and health, where a model needs to track details across many steps rather than answer a single question and stop.
Active Memory and Active Focus, explained
Active Memory lets Muse Wonder carry relevant details forward across a conversation, so context doesn't reset once a session runs long or a user returns after a break. Active Focus is built for multi-step tasks: Meta says the model can keep working through a job, recover from errors it makes along the way, and pause to ask the user for clarification before taking an action with real consequences, rather than guessing and pushing ahead on its own.
Access, pricing and where it was built
Muse Wonder is available through Meta's Model API, priced at $4 per million input tokens and $20 per million output tokens, as well as through Muse Code, Meta's coding-focused product. Muse Wonder Full, the larger of the two models, requires a Muse Code subscription. Meta says both models were trained at Prometheus, a gigawatt-scale data center the company has built out for frontier AI training.
What's confirmed and what isn't
Confirmed: Meta announced Muse Wonder and Muse Wonder Full on October 12, 2026, with Active Memory and Active Focus as the headline features, Model API pricing of $4 per million input tokens and $20 per million output tokens, and training at Meta's Prometheus data center. Reported scores on the Artificial Analysis Intelligence Index place Muse Wonder in the low-to-mid 50s, though the exact figure varies slightly by source. Not confirmed: pricing for any consumer-facing Muse app subscription tied to Wonder Full, and a rollout timeline beyond Meta's own API and Muse Code.
Primary reporting from Meta's research blog, with independent coverage from outlets including Engadget and Android Authority. Meta has not published consumer subscription pricing for Muse Wonder Full; confirm current pricing and availability before relying on it.
Frequently asked questions
What is Muse Wonder, and what did Meta announce?
Muse Wonder and the larger Muse Wonder Full are AI models Meta introduced on October 12, 2026, built around two features called Active Memory and Active Focus rather than a single benchmark claim. Meta positions them for tasks like document work, coding, and work in finance, law, and health.
What do Active Memory and Active Focus actually do?
Active Memory lets the model carry relevant context forward across a conversation instead of losing it. Active Focus keeps the model working through multi-step tasks, lets it recover from its own errors, and has it check in with the user before taking an action with real consequences.
How can people access Muse Wonder and Muse Wonder Full?
Muse Wonder is available through Meta's Model API, priced at $4 per million input tokens and $20 per million output tokens, and through Muse Code. Muse Wonder Full, the larger model, requires a Muse Code subscription.
How does Muse Wonder perform on benchmarks, and where was it trained?
Meta trained both models at its Prometheus data center, a gigawatt-scale facility built for frontier AI work. Reported scores on the Artificial Analysis Intelligence Index place Muse Wonder in the low-to-mid 50s, though the exact number varies slightly across sources.
What's confirmed and what isn't about pricing and availability?
Confirmed: the October 12, 2026 announcement, Active Memory and Active Focus, API pricing, and training at Prometheus. Not confirmed: pricing for any consumer Muse app subscription tied to Wonder Full, and a rollout timeline beyond Meta's own API and Muse Code.