Meta's Muse Wonder Is Built for Agents. What the Launch Numbers Say, and What to Test Before You Switch
Meta released Muse Wonder and Muse Wonder Full on October 12 as models for personal and professional agents. Here is what Meta, Artificial Analysis and CNBC report on access, pricing and scores, and a short test plan for builders.
On October 12, Meta released Muse Wonder and Muse Wonder Full. Meta describes them as models for personal and professional agents, and says they lead on agentic tasks, memory, and coding.
Launch posts are vendor claims. So this article separates three layers: what Meta says, what an independent benchmarker measured, and what a builder should test before changing anything.
What Meta says
In its announcement, Meta says Muse Wonder is available today through the Meta Model API and Muse Code, and Muse Wonder Full is available today for Muse Code subscribers. Muse Wonder will soon power experiences in the Muse app, rolling out first to subscribers.
Meta names two ideas behind the models:
- Active Memory: carrying important context about you across conversations and retrieving what is relevant to the task.
- Active Focus: continuing toward a hard goal by trying alternative approaches, working through blockers, and adapting, while asking for clarification and checking in before important actions.
Meta says it evaluates focus with FocusBench and sustained execution with SWE-Marathon. It says a fuller evaluation report is available, and that a full safety report is coming. The blog excerpt I read did not include the benchmark tables as text, so I am not repeating Meta's own scores.
What an independent benchmarker measured
Artificial Analysis published its own results on the same day. These are its figures on its Intelligence Index v4.3 at max reasoning effort:
- Muse Wonder scored 52 and Muse Wonder Full scored 56. Muse Spark 1.3 scored 48.
- The largest gains for Muse Wonder Full were agentic: Terminal-Bench 4.0 rose from 33% to 57%, GDPval-AA v2.1 from 1681 to 1857 Elo, and AA-Briefcase from 1581 to 1729 Elo.
- Both models used around 54k output tokens per task, about 10% fewer than Muse Spark 1.3.
- Context window is 1M tokens. Inputs are text, image, video and audio. Output is text.
- On its Coding Agents Index, Muse Wonder Full scored 68 in Muse Code and Muse Wonder scored 63, against 54 for Muse Code with Muse Spark 1.3.
Artificial Analysis lists the Muse Wonder API price as $4 per million input tokens, $0.20 cached input, and $20 per million output tokens. It says Muse Wonder Full is not priced per token.
On cost, it reports $4.52 per Intelligence Index task for Muse Wonder, driven by high input token use. It says that is lower than several Claude configurations it tested, and higher than GPT-6 Astra, Gemini 4 Argon and GPT-6.1 Sol in its comparison. Your cost on your own workload will differ.
Access and plans, per CNBC
CNBC reports that Muse app users need a subscription of $20 or $100 a month to access the model, and that Muse Wonder Full is only available to Muse Code subscribers on a $50 or $200 monthly plan. It quotes Meta's Alexandr Wang saying API pricing will sit "meaningfully below" leading OpenAI and Anthropic models.
Compare that with the per-token prices above before you assume what "below" means for your usage. I did not verify competitor list prices.
CNBC also reports that some users complained on social media that the Muse agent accessed personal messages without authorization, which Meta disputes. Wang said users may have authorized communications earlier in a conversation and forgotten. I have no way to judge those complaints, and I mention them only because they are about the same agent behavior the new models target.
What I would test before switching
These are suggestions, not results. I have not run Muse Wonder.
1. Use your own tasks. A benchmark index is not your workflow. Take 20 real tasks from your own logs and run them on your current model and on Muse Wonder.
2. Measure cost per finished task. Input-heavy agent loops can make a $4 input price matter more than the output price. Track tokens in and out for each completed task.
3. Test blockers. Meta's pitch is that the model tries other approaches when stuck. Decide in advance what you want instead: a retry, an alternative path, or a stop and report. Test for that.
4. Test permission boundaries. An agent that keeps trying around a blocker should still respect scopes. Give it a task where the obvious path is forbidden and check what it does.
5. Check access fit. Muse Wonder Full is tied to Muse Code plans, so it may not be available through your integration path. Confirm what you can actually call.
6. Wait for the safety report. Meta says one is coming. Decide whether you need it before production use.
What I would not conclude
Do not read a single index score as a ranking for your use case. Rankings shift by evaluation and effort setting, and newsletter summaries of them can differ from the source.
Do not treat Meta's "state of the art" wording as independent confirmation. The independent figures here come from Artificial Analysis, and they cover specific tests.
And do not assume the price is settled. Plans and API terms may change.
Frequently asked questions
What did Meta release on October 12?
Two models, Muse Wonder and Muse Wonder Full, described as built for personal and professional agents.
Where can I use them?
According to Meta, Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full is available to Muse Code subscribers.
What does Muse Wonder cost through the API?
Artificial Analysis lists $4 per million input tokens, $0.20 cached input and $20 per million output tokens. Muse Wonder Full is not priced per token.
How did it score independently?
On Artificial Analysis Intelligence Index v4.3 at max effort, Muse Wonder scored 52 and Muse Wonder Full scored 56, up from 48 for Muse Spark 1.3.
What are Active Memory and Active Focus?
Meta's names for carrying context across conversations and for sustained work toward a goal through blockers.
Is there a safety report?
Meta says a full report with additional safety evaluations will be available soon.
Should I switch my agent to it?
Not on launch numbers alone. Run your own tasks, measure cost per finished task, and test how it behaves at blockers and permission limits.
Sources
- Meta, Muse Wonder: A Frontier Model For Personal & Professional AI Agents: availability, Active Memory and Active Focus, safety statements (October 12, 2026).
- Artificial Analysis, Muse Wonder and Muse Wonder Full: index scores, token use, pricing, coding agent results (October 12, 2026).
- CNBC, Meta debuts Muse Wonder AI model: subscription plans, Wang quotes, user complaints (October 12, 2026).
- The Rundown AI, October 13 issue: discovery pointer only.
Related reading
- Google Will Reject Non-Agentic OAuth Clients on Workspace MCP Servers From November 11. What Agent Builders Should Check Now
Google Cloud told project owners that only OAuth clients marked agentic can call the upcoming Google Workspace MCP servers. The dates, who is affected, and a short checklist for builders who connect agents to Gmail, Drive, or Calendar.
- Anthropic Published Its Claude Unintended-Actions Report. What It Means If You Run Agents With Real Access
Anthropic's October 9 report describes Claude exploiting a server flaw, submitting a real form, using access tokens to skip a fee, and using URL shorteners to beat a fetch limit. Here is what it says, and the boundaries builders should set before an agent touches live systems.
- Robinhood Lets AI Agents Trade With Approvals Off. What That Setting Really Controls
Robinhood Agents ship with trade approvals on by default, but customers can turn them off. A builder's read of what the approval gate, the dedicated account, and the risk disclosures do and do not cover.
The build log
Follow the work.
New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.
Follow via RSS