Meta wants your code. Badly enough to hand you a discount of more than 20x to get it. On 5 August it launched Muse Code, a terminal coding agent, and the pricing tells you exactly what the game is.
Quick translation for anyone not neck-deep in dev tools: a terminal coding agent is an AI that lives in your command line and actually writes, runs and fixes code across a whole project, rather than just finishing your sentences like autocomplete. Muse Code runs on Muse Spark 1.2, Meta’s newest model, and it can keep several “background agents” grinding away on separate tasks while you go and do something else.
The two-tier bribe
Standard pay-as-you-go pricing is $1.25 per million input tokens and $4.25 per million output. Roughly competitive, nothing to see. Then there is the “Contributor” tier: $0.10 input, $0.20 output, with cached input at a rounding-error $0.002. That is over 12x cheaper on the way in and more than 20x cheaper on the way out.
The catch is the whole point. To get the cheap rate you agree to let Meta use your prompts and the model’s completions, meaning your real code and its real answers, to train future models. And Muse Code defaults to the Contributor tier the moment you install it. You have to go out of your way to switch to Standard to keep your work to yourself.
Who actually pays here
Read the trade for what it is. Meta, a company sitting on more cash than most economies, is offering a 20x discount to indie devs and cash-strapped startups in return for the one thing those people own outright: their code, their debugging trajectories, the messy real-world data that money genuinely cannot buy at this scale. The people most likely to click “yes” are the ones least able to afford full price. That is not a coincidence, that is the design. The cheaper it feels, the more you are the product.
The land grab
The strategy has a name in the trade: land and expand. Get a huge base of developers hooked at a loss-leading price, hoover up their agentic data, then use that data to build models good enough to charge properly for later. OpenAI and Anthropic charge for Codex and Claude Code and largely leave your code alone. Meta is underpricing both by an order of magnitude and asking for the code instead of the cash. Whether that is a bargain or a Faustian expense account depends entirely on what you are building.
Is the thing any good?
Middling-to-decent. Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index, up three points from 1.1, which lands it level with Grok 4.5 and GPT-5.5 and comfortably behind the frontier (Claude Opus 5, GPT-5.6 Sol, Kimi K3). So you are not getting best-in-class in exchange for your data. You are getting good-enough, at a price designed to make good-enough irresistible.
If you are a hobbyist building a weekend game, fine, feed Meta your photon-sphere simulation, who cares. If you are writing anything proprietary, read the tier toggle before you type a single prompt.
Did you know: the “tokens” you are billed for are chunks of text of roughly four characters each, so a single million-token bill is about 750,000 words, or the King James Bible and most of a sequel.
Related on Top Tool Stack: Claude Cowork goes cloud · Thinking Machines’ Inkling-Small