How to Use Muse Wonder: API, Muse Code and a First Task

Short answer: today there are two ways to use Muse Wonder, the Meta Model API and Muse Code, Meta’s coding agent for the terminal. It is not in the Muse app yet. Meta released Muse Wonder and Muse Wonder Full on 12 October 2026 and says Wonder is available to developers through the API and Muse Code, with Wonder Full for Muse Code subscribers.

This guide walks through both routes, works out what a typical task costs, and gives you one first task that shows quickly whether the model is any use for your work. Everything here is from Meta’s own pages as of 15 October 2026. Prices are in US dollars.

First, which Muse is which

Meta now has four things with nearly the same name, and most of the confusion starts here. Muse is the personal agent app that launched in the US in September. Muse Spark is the model family behind Meta AI. Muse Code is the coding agent. Muse Wonder is a new, larger model, and Muse Wonder Full is a stronger version of it. The older family is explained in what is Muse Spark, and if Spark is what you actually have, the Muse Spark tutorial is the better guide.

The point that matters for this guide: a model is not something you install. You either call it through an API or use a tool that has it built in. Those are the two routes below.

Route 1: the Meta Model API

The API is the direct route, and the only one where you pay per use rather than by subscription. Meta’s quickstart asks for a Model API account and one of three tools: curl, Python 3.9 or later, or Node.js 18 or later, with the openai package for the last two.

1. 01Sign in to the Model API dashboard and open API keys.
2. 02Choose Create API key and copy it somewhere safe straight away.
3. 03Store it as an environment variable called MODEL_API_KEY rather than pasting it into code. On macOS and Linux that is one export line in your shell.
4. 04Send a request to the responses endpoint, api.meta.ai/v1/responses, with the model set to muse-wonder-1.0. The quickstart’s own example uses a Muse Spark model, so change that one line.
5. 05If you already use an OpenAI-compatible client, you do not need to learn anything new. The Muse Wonder model page says to point that client at the Meta Model API. Note that the OpenAI libraries will not read MODEL_API_KEY on their own, so pass the key in explicitly.

The model page also describes search grounding: add one tool to the request and Wonder can base its answer on live web results. Leave that off for your first test. You want to see what the model does with material you gave it, not what it found.

What one task costs

Meta’s price list puts Muse Wonder at $4 per million input tokens, $0.20 per million cached input tokens and $20 per million output tokens, with a 1 million token context window.

A worked example. You send a project report of about 20,000 tokens, which is a long document, and get back a 2,000 token summary with an action list. The input costs $0.08 and the output $0.04, so 12 cents in all. Costs grow when an agent works for a long time, writes many steps of its own and resends the whole context each time. Output costs five times as much per token as input, so the first lever on the bill is the length of the answers you ask for.

How that price compares with Google’s newest model is in Muse Wonder vs Gemini.

Route 2: Muse Code

Muse Code needs no API key set up by hand. Per the quickstart, you install the command-line tool on macOS or Linux, run muse, and on first launch either sign in through your browser or paste a key. The quickstart lists no Windows installer.

Muse Wonder Full is only available here, not on the API. The model page names two plans that include it: Power Usage at $50 a month and Maximum Usage at $200 a month. Muse Code’s plans page starts at $5 a month. How Muse Code compares with the obvious alternative, and what its cheaper plans trade on data, is in Muse Code vs Claude Code.

Which route suits you? If you write code and already live in a terminal, Muse Code gets you to a result faster. If you want Wonder inside your own tool, such as an internal form or a spreadsheet script, you need the API. If you do not write code at all, wait for the app.

What about the Muse app?

Meta says Wonder will power more features in the Muse app soon, starting with subscribers. It gives no date. Neither the launch post, the model page nor the quickstart lists countries, so if you are outside the US the reliable test is your own account. Using a VPN or someone else’s account to get around that is not something we recommend: it can break the terms of use and leaves you with no support if something goes wrong.

What Muse Wonder is built for

According to Meta’s research post, Wonder is built for long, multi-step work. It plans, hands parts of a task to sub-agents, calls tools in parallel and is meant to recover from its own mistakes. Meta also highlights Active Memory, which carries context about you from one conversation to the next. TechCrunch describes it as Meta’s first frontier model.

Those are Meta’s claims, not independent tests. In practice it means Wonder earns its price on tasks with many steps. For a two-line email it is expensive and more than you need. What an agent is, and when one helps, is covered with real cases in AI agent examples.

A first task that tells you something

Start with material where you already know the right answer. Give Wonder a report you wrote yourself and ask for a summary. You will see straight away what it leaves out, what it adds that was never there, and how long its answers run. Only then is it worth giving it material you have not read.

Agent models rarely fail for lack of ability and often fail on a vague brief. Instead of "Make me a plan for this project", try something like this:

Here is our project report. List the open items with an owner and a deadline for each. Where the report names no owner, write "unassigned" rather than guessing one. One page at most.

Three things do the work: a clear output format, an instruction for missing information, and a length limit that also caps the cost. More patterns like this are in how to write a prompt that works first time, and they apply to Wonder unchanged.

Before you use it for real work

• —Decide what data can go in. Wonder remembers things across conversations, and Meta’s full safety report is still "available soon". Read is Muse Wonder safe before client or staff data goes anywhere near it.
• —Decide who checks the output. A model that takes ten steps on its own can also be wrong for ten steps. How to check an AI answer when you are not the expert is the routine to use.
• —Set a budget. Decide what a test may cost before you leave an agent running, and watch usage in the dashboard.
• —Expect change. The model is days old. Behaviour, limits and prices can move, so do not build anything critical on it yet.

New models now arrive every few weeks. What lasts is the skill underneath: writing a clear brief, checking what comes back, and knowing where your data goes. Coursium teaches people to use AI at work, on their phone, in short lessons on exactly that. You can see how Coursium works or get it on the App Store.