With Muse Code, only the tool is free. The AI behind it costs money from the very first prompt — and it only gets cheaper if you hand Meta your code.
In early August, Meta launched its first in-house coding agent: Muse Code, currently still in beta. What it does is quickly told. How Meta bills for it is the interesting part.
What Muse Code actually does
Muse Code is what people call a terminal agent. Meaning: no chat window in your browser, but a program running in the command line — that black window where developers type commands. You describe what should happen, and the AI works its way through an entire software project on its own: planning the change, writing the code, then checking the result.
Meta isn't first here. Anthropic has Claude Code, OpenAI has Codex, Google has its own counterpart. Muse Code is Meta's answer, built by its in-house Superintelligence Labs.
One detail is genuinely well solved: Muse Code logs every step — every model call, every file change, every approval. If the program crashes mid-job, it picks up exactly where it stopped. Anyone who has ever lost half an hour of AI work to a crash knows why that's worth more than any benchmark point.
Speaking of benchmarks: on the common coding test, Meta's model currently sits behind the top models from Anthropic and OpenAI. Not a disaster, but not a sensation either.
What's free — and what isn't?
The tool itself costs nothing. No subscription, no per-seat price, no catch.
The catch is the model underneath. Muse Code needs Muse Spark 1.2, Meta's AI trained specifically for programming — and that is billed per token. A token is roughly a syllable; a longer session racks up tens of thousands of them fast. At the current rate, the standard tier sits in the low single-digit dollars per million input tokens, and somewhat higher for what the AI writes back.
And here's the part worth knowing up front: there is no free credit, no trial, no free tier. Without a payment method you don't get as far as the first prompt.
To be fair: Claude Code and Codex aren't free either, and the free tiers of ChatGPT, Claude or Gemini cover the chat window in your browser — not agents of this class. For tools that work on their own for hours, free is the exception, not the rule. Compute costs providers real money. So the gap is smaller than it first sounds. Only: with Muse Code the meter starts without a run-up.
On top of that: Muse Code decides for itself how many tokens to burn. It spins up several helper agents in parallel, each making its own requests. There is no spending cap. Anyone who turns it loose on a large project unsupervised should keep an eye on the meter.
The contributor tier: cheaper if Meta gets to read along
This is where it gets interesting. Meta offers a second tier that pushes token prices down dramatically — roughly to a tenth or even a twentieth of the normal rate.
Important: cheaper, not free. The bill still arrives, it's just a lot smaller. And the discount is stated openly in the terms: Meta gets to train its future models on your inputs and the AI's answers. Your code becomes training material.
The cheap tier is also throttled far harder — you may only send a fraction of the requests per minute. Cheap, yes. Slow, also yes.
Is that a scandal? No. It's an honestly labelled trade, and Meta itself notes it suits experiments — not customer projects. That's exactly how to treat it: hobby project or throwaway code, fine. Your employer's source code or anything with customer data in it? Absolutely not.
One side note observers picked up on: open source goes unmentioned in Meta's announcement this time. For a company that spent years positioning itself as the open counterweight to OpenAI, that's a remarkable gap.
Is this for me?
Honest answer: only if you write code. Muse Code is not a beginner's tool. It runs on macOS and Linux; on Windows only via a detour that already requires prior knowledge.
If you've never opened a command line, this isn't your entry point — there are friendlier tools for that.
What you can take away anyway: the pattern behind this pricing model is everywhere. With AI services, "free" almost never means free of charge — and where it gets cheaper, the discount is usually paid for. With waiting time, with a weaker model, with your data. With Muse Code, at least it says so on the label.
The question is rarely whether you pay. It's how much — and what you hand over on top to push the price down.