Meta’s New Coding AI Is 21x Cheaper. You Pay With Your Code

TL;DR: On 5 August 2026, Meta launched Muse Code, its first AI coding agent — a direct rival to Claude Code and OpenAI’s Codex. It has two prices. The cheap one is up to 21x cheaper, and the reason is printed right on the tin: on that tier you give Meta permission to train its future models on your prompts and your code. It’s also terminal-only, beta, and doesn’t run on Windows — so for most non-developers there’s nothing to do today. But the pricing model is worth understanding, because “much cheaper if you hand over your data” is a trade you’ll be offered again.

What happened

Meta launched Muse Code on 5 August 2026, powered by a new model called Muse Spark 1.2. It’s an AI agent that plans changes, writes code, checks its own work, and can run background agents that keep going throughout a session — including splitting a big job across several sub-agents at once.

The practical details matter more than the capabilities here:

• It’s terminal-only. As 9to5Mac notes, “unlike ChatGPT Codex and Claude Code, there’s currently no dedicated app interface” — you install it with a command line and drive it by typing commands.
• macOS and Linux only. No Windows support at launch.
• It’s in beta.
• There’s no subscription. You pay per use, through Meta’s model API — which is where it gets interesting.

The two prices, and what the cheap one costs you

Meta is offering the model on two tiers, and the gap between them is enormous:

	Standard	Contributor
Input (per million tokens)	$1.25	$0.10
Output (per million tokens)	$4.25	$0.20
Meta trains on your data?	No	Yes
Rate limit	3,000 requests/min	60 requests/min

The Contributor tier is roughly 12x cheaper on input and 21x cheaper on output. In exchange, as Implicator reported, you explicitly grant Meta permission to train future models on your prompts and your code completions. Meta’s AI chief framed it as developers choosing to “opt-in to help improve the model,” and it is a genuine opt-in — nobody is hiding it. The Standard tier carries a commitment not to train on what you send.

Two details worth noticing that get less attention: the cheap tier is also rate-limited far more tightly (60 requests a minute against 3,000), and Meta is only “beginning to accept requests” for zero-data-retention arrangements — so the no-training promise on Standard isn’t quite the same thing as “we don’t keep it.”

Why it matters if you don’t code

Let’s be straight about the tool itself: there’s nothing here for most non-developers today. A terminal-only beta that doesn’t run on Windows, with no app and per-token billing, is aimed squarely at professional developers. If you build with Lovable, Bolt, Replit or Claude Code’s app, your options haven’t changed this week. This is the same honest verdict we reached about Kimi K3’s open weights — a big story that simply isn’t yours to act on yet.

What is worth carrying away is the pricing structure, because it makes something explicit that’s usually buried:

• “Cheaper” often means you’re paying with data instead of money. Meta has done the unusual thing of putting a precise number on it — your prompts and code are worth about a 21x discount. Most tools never spell that out. Now you have a reference price for the trade.
• The thing being traded here is your actual work. Not usage statistics — the prompts you write and the code that comes back. If you were building something with client data, business logic, or an idea you’d rather keep to yourself, that’s what’s on the table.
• It’s the same question we keep landing on, from an AI coding tool that was caught uploading users’ code to deciding how much access to give agents over your email and files. The difference this time is that it’s consensual and clearly priced — which honestly makes it a healthier version of the trade than most.

There’s also a straightforward comparison worth making. Per-token billing across two tiers with different data terms is exactly the kind of decision non-developers find hard. A flat monthly subscription — what Claude Code and most beginner-friendly tools use — is easier to reason about, and you can’t accidentally agree to anything by picking the cheaper number.

What to do

• Nothing, if you’re a non-developer. It doesn’t run on Windows, has no app, and is in beta. Your current tools are unaffected.
• If you’re technical and tempted by the cheap tier, ask one question first: would you be comfortable with this code and these prompts being used to train a model? If it’s a personal side project, quite possibly yes. If it’s client work or anything commercially sensitive, pay the standard rate.
• Everyone: use this as a reference point. When any AI tool offers you a notably cheaper plan, look for what it’s asking for in return. It won’t always be labelled as clearly as Meta has labelled it.

Who should care (and who shouldn’t)

• Building with Lovable, Bolt, Replit or Claude Code’s app: nothing to do — this doesn’t reach your workflow.
• Comfortable in a terminal, on a Mac or Linux, and cost-sensitive: genuinely worth a look, with the data question answered honestly first.
• Working on anything commercially sensitive: stay on Standard, or stay where you are.
• Not sure which tool suits you: the quiz matches you on what you’re building — a new developer-focused terminal agent doesn’t change that answer.

Our take

We rate the transparency here, genuinely. Plenty of companies quietly train on user data; Meta has instead put two prices side by side and let you choose, which at least lets people make an informed decision rather than an accidental one. That’s better behaviour than the industry norm.

We’d still be cautious about the Contributor tier for anything that isn’t a hobby project — “cheap” stops being cheap if the code you hand over is the thing that makes your business yours. And for this site’s readers specifically, the practical answer is easy: a terminal-only beta with no Windows support isn’t a tool you need. Watch the pricing idea, not the product.

Not sure which AI tool actually fits how you build? Take the 60-second Vibe Coding Tool Finder quiz →

FAQ