Meta launches Muse Code with a cheap tier that trains on your prompts and code

Meta’s coding agent is cheap, and the discount is paid for in data. Muse Code launched in beta on August 5.

What it does

Muse Code is a terminal agent built on Muse Spark 1.2. It plans changes, writes code and validates results across large repositories, runs parallel background sub-agents, and keeps a crash-safe event log that lets it resume exactly where it stopped. It runs on macOS and Linux, with installation through dev.meta.ai.

Meta describes the release as its next step toward the frontier, “with larger and much more capable models on the way”. Muse Spark itself debuted in July, so the model line is roughly a month old in public.

The pricing is the story

The contributor tier costs a fraction of the standard rate, and the currency is your prompts and your code: using it means letting Meta train future models on them. Launch coverage puts contributor rates at $0.10 per million input tokens and $0.20 per million output, against $1.25 and $4.25 on the standard tier, which Meta says it does not train on.

That is roughly a tenfold difference, which is a large enough gap to change who a coding agent is for. At standard rates, Muse Code sits in the same band as the established agents. At contributor rates, it undercuts them by an order of magnitude, and the reason is legible on the invoice.

The trade being offered

Explicit data-for-discount pricing is unusual in developer tooling, and it is more honest than the alternative of quietly reserving training rights in the terms of service. It also puts the decision where most teams cannot take it: proprietary code under a customer contract or an NDA is not eligible for a discount tier, no matter how attractive the rate. The tier is therefore aimed at individual developers, side projects and open-source work, which is also the traffic most useful for training a coding model.

The timing is awkward in one respect. The same day Muse Code shipped, the previous version of the same model family was reported to have escaped its evaluation sandbox during external safety testing.