Meta Muse Wonder: Active Memory, Active Focus and What It Costs
Executive Summary
- What shipped: On October 12, Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full, Meta's first frontier-class models, built for personal and professional agents.
- The two features: Active Memory carries context about you across conversations; Active Focus keeps an agent working toward a goal when it hits blockers, and checks in before important actions.
- Where it ranks: 52 on the Artificial Analysis Intelligence Index (Wonder Full: 56), behind Claude Opus 5.5 at 58. Its standout is long-running coding (SWE-Marathon), on Meta's own chart.
- Price and access: $4 per million input tokens and $20 per million output in the Meta Model API, the same as Claude Opus 5.5. Wonder Full is only in Muse Code subscriptions. No open weights.
I made a 45-second video of this launch on X; this post adds the numbers, the limits and what to check before you switch.
Active Memory: context that survives the chat
Meta describes Active Memory as carrying important context about you across conversations and pulling in the details that matter for the current task. In practice that means you stop repeating yourself: preferences, ongoing projects and constraints from earlier chats are recalled when a new request needs them.
Illustration: details saved on Monday are recalled in a new chat on Thursday
Meta's model card evaluates this with MemoryBench, an internal test of recall across long multi-turn conversations of 64K to 100K tokens, with distractor turns and updates that change earlier facts. It is Meta's own benchmark, so treat the claim as promising rather than proven until independent memory tests exist.
Active Focus: agents that re-plan instead of giving up
Active Focus is about long tasks. When a step fails, the agent tries an alternative, adapts to new information, asks for clarification, and checks in before important actions such as paying or sending something on your behalf. Meta measures it with FocusBench, another internal evaluation of real-world agent tasks in areas like travel, housing, IT and finance.
Illustration: a step fails, the agent adds a new plan, and asks before paying
Benchmarks: close to the top, not first
On the independent Artificial Analysis Intelligence Index, as reported by VentureBeat:
- Claude Opus 5.5: 58
- Muse Wonder Full: 56
- Claude Sonnet 5.5: 56
- GPT-6 Astra: 53
- Gemini 4 Argon: 53
- Muse Wonder: 52
Artificial Analysis Intelligence Index: Muse Wonder Full at 56 and Muse Wonder at 52, behind Claude Opus 5.5 at 58
Where Meta's chart shows a clear lead is long-running software work. On SWE-Marathon, Muse Wonder Full scores 57.5 and Muse Wonder 53.1, against 46.3 for Claude Opus 5.5 and 42.5 for GPT-6 Astra. Meta's model card notes that the competitors' results used different agent harnesses (Codex and Claude Code), so this is not a controlled comparison.
SWE-Marathon on Meta's chart: Muse Wonder Full 57.5, Muse Wonder 53.1, Claude Opus 5.5 46.3, GPT-6 Astra 42.5
Price, limits and access
According to the model card, TechCrunch and VentureBeat:
- Model:
muse-wonder-1.0in the Meta Model API; text, image, audio and video in, text out; about 1M tokens of context. - Price: $4 input and $20 output per million tokens, $0.20 for cached input, $2.50 per 1,000 web searches. Artificial Analysis puts the cost of its average index task at $4.52, between GPT-6 Astra ($3.26) and Opus 5.5 ($5.98).
- Limits: 3,000 requests and 4 million tokens per minute.
- Wonder Full: not in the API. It is available through Muse Code subscriptions (Power at $50 a month and Max at $200, per VentureBeat).
- Weights: not released.
Check this before you switch
Per VentureBeat's reading of Meta's terms, the pay-as-you-go API runs in a standard mode where your content is not used for training, but Muse Code subscriptions default to a "Contributor mode" in which Meta may use prompts, code, files and outputs to improve future models. Turning it off only applies to future submissions. If you plan to use Wonder Full on a work codebase, change that setting first.
A practical checklist:
- Run your own tasks, not just the leaderboard: the gap to Opus 5.5 is small on some tests and large on others.
- If you build agents, test long tasks with real blockers; that is where Active Focus should show.
- Compare cost per finished task, not per token: Wonder has the same per-token price as Opus 5.5 but cost less per Artificial Analysis task ($4.52 vs $5.98).
- Check data settings before connecting private repos.
Sources
Comments
Share your thoughts and join the conversation
