How to Use Muse Spark: The Meta AI App, Muse and the API

Muse Spark is not an app you download. It is a model, and you reach it through one of three doors. Which door you pick decides almost everything about the experience: a chat window, an agent that does tasks for you, or a developer tool. Most confusion about how to use Muse Spark comes from mixing those up.

If you want the background first — who built it, and why the name keeps turning up in two places — what Muse Spark is covers that. This post is the practical half: where to go, which mode to choose, and one real work task done end to end.

The three doors, in one list

• —The Meta AI app and the meta.ai website. This is a chatbot. You ask, it answers. Meta says both have run on Muse Spark since its launch on 8 April 2026.
• —Muse, the agent app. This one does tasks rather than just answering: emails, bookings, forms, plans. It runs on the same model family.
• —The Meta Model API and Muse Code. These are for developers building on the model or letting it work inside a code base.

If your goal is to think through a piece of work, start with the first door. That is where most people should begin, and it is where the rest of this post spends its time.

Door one: the Meta AI app and meta.ai

You need a Facebook or Instagram account to sign in to meta.ai, as Simon Willison found in his hands-on test on launch day. If you have neither, that is the first hurdle, and it is worth deciding whether you want your work questions tied to that account before you start.

Once you are in, the chat can do more than write text. The same hands-on review lists web search and page opening, a Python sandbox for working with data, finding and counting objects in images, image generation, and small HTML or SVG outputs. It can also search public Instagram, Threads and Facebook posts, and link a calendar or email account if you choose to.

On "try Muse Spark free": the chat is the lowest-friction way in, since it only asks for a login. The separate Muse agent app has its own free tier, covered below, and it works differently.

Instant vs Thinking, and what Contemplating is

The chat offers two modes. Instant answers straight away. Thinking takes longer and works through the problem before replying. Meta also describes a third, Contemplating, where several agents reason in parallel — and says it is rolling out gradually, so you may not see it yet.

A simple rule for choosing:

• —Instant for rewording, short summaries, quick definitions and first drafts you will edit anyway.
• —Thinking for anything with steps: comparing options, working with numbers, planning, or a question where a wrong answer costs you something.
• —Contemplating, if you have it, for hard problems where you would happily wait. Do not use it as a default; slower is not automatically better for a two-line email.

You will see Contemplating quoted with a headline number: Meta reports 58% on Humanity’s Last Exam in that mode. That is a score on a test set. It tells you the model is capable, not that it will be right about your supplier contract. Meta itself, in the same launch post, named gaps in long-horizon agent work and coding.

Door two: Muse, the agent app

Muse launched on 8 September 2026, rolling out in the US on iOS, Android and muse.ai. It is built to do things: send emails, book travel, fill out forms, make plans, and keep working after you close the app. You choose which services it connects to and how much access each gets, and Meta says it asks for approval before sending an email or making a purchase.

There is a free tier with a usage meter, plus paid tiers at $20 and $100 a month, as of September 2026. The same report notes that a payment card is needed to start, even on the free tier.

An agent that can send and buy is a different kind of tool from a chatbot, so it deserves its own setup. Setting up and using Muse, Meta’s agent app walks through that. Before you connect your work email, read what Muse can see and how its permissions work.

Door three: the Muse Spark API and Muse Code

Developers reach the model through the Meta Model API, which opened as a public preview in July with a 1 million token context window, tool use, and support for MCP servers. The current version, Muse Spark 1.3, uses about 20% fewer tool calls and about 25% fewer tokens than the version before, according to Meta.

Standard API pricing is $1.25 per million input tokens and $4.25 per million output tokens. There is also a cheaper contributor tier, but it requires letting Meta use your prompts for training. For company data, that trade matters more than the discount.

Muse Code is a coding agent that runs in your terminal. You install it with a one-line script, sign in through the browser, and run it inside a project folder, as described on Meta’s developer blog. It splits work across parallel subagents and has a plan step that needs your approval before anything runs. If you are choosing between coding agents, Muse Code vs Claude Code compares the two.

A worked example in Thinking mode

Here is a task that comes up in most offices: three quotes for the same service, each priced a different way, and a manager who wants a recommendation by Friday.

Switch the chat to Thinking. Then paste the numbers as plain text, with no names or details you would not want stored. A request that works looks like this:

I have three quotes for office cleaning. Supplier A: £420 a month, 12-month contract. Supplier B: £95 a week, no contract, £150 setup fee. Supplier C: £4,600 a year paid upfront, includes two deep cleans. Work out the total cost of each over 12 months. Show the arithmetic line by line. List every assumption you make. Then write a four-sentence recommendation for my manager, who cares about cost and about not being locked in.

Notice what that request does. It names the job, the reader, the shape of the answer, and asks for the working. Those are the same four things in how to write a prompt that works on the first try, and they matter as much here as anywhere.

A good answer will say that Supplier B depends on how many weeks it assumes in a year, will total each option, and will flag that Supplier C’s deep cleans only count if you would otherwise pay for them. A weak answer will skip one of those without telling you.

How to check the result

The answer will read calmly and confidently whether it is right or not. That is a known property of these models, with its own body of research, not a quirk of one product. So check it in this order:

1. 01Redo one sum yourself. Supplier B is the easiest to get wrong: 52 weeks at £95, plus the fee. If that number is off, distrust the rest.
2. 02Read the assumptions list. Anything you did not tell it, it guessed. Correct the guesses that matter.
3. 03Check the recommendation against the numbers above it. Models sometimes write a conclusion that the working does not support.
4. 04If it got something wrong, do not argue in the same chat. Fix the request and start a fresh one.

Arithmetic is one of the places AI is actually bad at when it predicts numbers instead of calculating them, which is why asking for line-by-line working earns its place in the prompt. For claims outside your own field, how to check an AI answer when you are not the expert sets out a routine.

The cleaning quotes are a stand-in. The same pattern works for comparing software plans, travel options or job offers: give the numbers, ask for the working and the assumptions, and check one line yourself.

What to do on Monday

1. 01Decide which door you need. For thinking through work, that is the Meta AI app or meta.ai, not the agent app.
2. 02Pick one real task from last week that had numbers or steps in it. Not a test question; something you actually did.
3. 03Run it in Thinking mode with the job, the reader, the format and a request for the working.
4. 04Check one number and read the assumptions before you use any of it.
5. 05Write down what it got wrong. After a few tasks you will know where this model helps you and where it does not.

If a task comes back every week, it may be worth more than a chat. Finding the repetitive part of your work is the step before handing anything to an agent like Muse.

Benchmarks and launch posts will keep changing. The habits above will not: pick the right mode, ask for the working, and check before you send. Coursium is an iPhone app that teaches people to use AI at work, with practice tasks built around exactly that kind of checking. Stay ahead of AI by learning the tools on your phone.