Meta has released Muse Wonder, and it’s the company’s most serious attempt yet at a frontier model for AI agents. Instead of answering one prompt at a time, it’s built to remember useful context, stay on a task when things get hard, write code and work across apps. In this guide I’ll explain what Muse Wonder and Muse Wonder Full are, what Meta says they can do, what they cost, and how I got Muse Wonder running in the terminal, including the setup mistake that cost me time.
Prefer to watch? See the full walkthrough on YouTube. I build the whole thing on screen.
What Muse Wonder is
Meta announced the Muse Wonder series on October 12, 2026. There are two models:
- Muse Wonder: available to developers through the Meta Model API and Muse Code, Meta’s coding agent.
- Muse Wonder Full: the bigger model, available only to Muse Code subscribers on the higher plans. It isn’t served through the API.
Both build on Muse Spark, Meta’s earlier lightweight model. According to the model card, they accept text, image, video and audio input, produce text output, and have a 1M-token context window. Meta says Muse Wonder will also power its Muse personal agent, rolling out to subscribers first.
Active Memory and Active Focus
Meta highlights two capabilities:
- Active Memory: carrying what matters to you across conversations and bringing back the details that are relevant to the current task.
- Active Focus: continuing to work toward a goal when a task gets hard, trying other approaches, recovering from errors, and checking in before doing anything important.
Those are the two things that usually break agents in practice. An agent that forgets your preferences or gives up after the first error isn’t much use for real work, so this is the right problem to aim at. Whether it holds up depends on your own tasks, which is why I’d test it before trusting it.
How it compares, according to Meta
Meta’s launch page shows Muse Wonder and Muse Wonder Full scoring at or near the top of several benchmarks, including the Artificial Analysis Intelligence Index, agentic automation and coding tests, and finance and legal agent tasks. TechCrunch’s write-up summarizes it as comparable to Claude Fable 5.1 and GPT-6 Astra, based on Meta’s own testing.
Keep that last part in mind. These are vendor-reported numbers on benchmarks the vendor chose. They’re a useful signal that Meta is now competitive, which hasn’t always been true, but they’re not a substitute for running the model on your own work.
The demos on Meta’s page
The launch page includes a few interactive examples: a small 3D kitchen game you can play in the browser, a wellness-support demo, an agent adding events to a calendar, and one dropping pins on a map through browser automation. Browser use has historically been a weak spot for Meta’s models, so the calendar and map demos are the interesting ones if you build agents that work in web apps.
What Muse Wonder costs
At launch, Meta’s developer blog lists Muse Wonder on the Meta Model API at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens. That puts it in the same price range as other frontier models, not the budget tier.
Muse Code is sold as subscription plans, from an Everyday Usage plan up to a new Maximum Usage plan with 80 times the Everyday limit. Muse Wonder Full is included only on the Power Usage and Maximum Usage plans. When I recorded, the plans ranged from about $5 to $200 a month. Check Meta’s pricing page for current figures, because launch pricing changes.
How I set up Muse Wonder
There are a few ways in. Here’s what I tried, in order, and what I’d do if I started again.
Option 1: Use the API key in OpenCode, Claude Code or Codex
You can create an API key in Meta’s developer portal and add Muse Wonder as a model in a coding harness you already use. I tried OpenCode first and kept getting an invalid API key error. The portal showed tokens being used and a spend of under a cent, so the key worked, but the model still wasn’t usable inside OpenCode.
I asked Claude Code to debug the configuration, which is how I fix most setup problems on Linux now. If you go this route, check that the provider name, base URL and model ID match Meta’s docs exactly. Those are the usual causes of a key that works in the portal but not in the tool.
Option 2: Install Muse Code and log in with the browser
The faster route was Meta’s own coding agent. Install Muse Code, launch it, trust the folder, and choose log in with browser instead of pasting an API key. The default model in my install was Muse Spark 1.3. Switch it with the /model command and pick Muse Wonder.
I asked it what it could do, and it listed the usual coding-agent jobs: reading and explaining code, finding and fixing bugs, adding features, writing tests and running builds. A simple test cost a fraction of a cent.
If you want a structured way to learn coding agents like this, the Claude Code and Codex courses in AI Income Lab walk through setup and everyday workflows step by step.
Option 3: A real test on a project
Next, I opened one of my own projects and asked Muse Wonder to run an audit and suggest fixes. It built an inventory of the repo’s health, ran checks and scanned for security issues. That run cost roughly a dollar in API usage, which is a reminder that audits on a full codebase add up faster than a quick question does. Set a spend limit in the portal before you point it at a large repo.
Who should try Muse Wonder
- Agent builders who need a model that keeps context across long tasks and handles browser and computer use.
- Developers who want another frontier coding model to compare against what they already use.
- Teams in law, finance or health exploring document-heavy agent tasks, which Meta calls out as a focus. Keep a human review step for anything that matters.
If you’re happy with your current model and don’t need the memory features, there’s no rush. Pick one real task, run it on both, and compare the output and the cost.
Common mistakes to avoid
- Fighting the API setup first. If you just want to try the model, Muse Code with browser login is quicker.
- Forgetting to switch models. Muse Code may default to Muse Spark. Use
/modelto pick Muse Wonder. - Running a full audit without a budget. Large codebases burn tokens. Set a spend limit and check the usage page.
- Treating launch benchmarks as proof. They’re Meta’s numbers. Test on your own tasks.
- Expecting Muse Wonder Full on the API. It’s subscription-only in Muse Code.
Got stuck on a setup step like this? AI Income Lab has a community where you can ask questions while you build.
Want help building this for your own business?
The video shows what’s possible. Inside AI Income Lab, the paid community from AI Automation Station, you get step-by-step courses, reusable templates, and a place to ask questions while you build. The courses walk you through builds like this one step by step. Over 2,900 members are already building with it, from $29 a month, and you can cancel anytime.
FAQ
What is the difference between Muse Wonder and Muse Wonder Full?
Muse Wonder is available on the Meta Model API and in Muse Code. Muse Wonder Full is Meta’s most capable model and is only available to Muse Code subscribers on the Power Usage and Maximum Usage plans.
How much does Muse Wonder cost?
At launch, $4 per million input tokens and $20 per million output tokens on the Meta Model API. Check Meta’s developer site for current pricing.
Can I use Muse Wonder in Claude Code, Codex or OpenCode?
You can add it with an API key, but I had configuration trouble in OpenCode. Muse Code with browser login worked straight away.
How big is the context window?
1M tokens, according to Meta’s model card.
Conclusion
Muse Wonder is the first Meta model I’d consider for real agent and coding work. The Active Memory and Active Focus features target the right problems, and the API price puts it in line with other frontier models. Start with Muse Code and browser login, give it one real task with a spend limit in place, and compare the result with the model you use today. Watch the video above to see the setup, mistakes included.
If you’d rather not figure this out alone, AI Automation Station has a paid community for it: AI Income Lab. It’s where the courses, templates and help that go with these builds live, with 2,900+ members, from $29 a month, cancel anytime.
Leave a Reply