Meta's coding agent has two prices: pay 18x more, or let it train on your repository

The discount is not flat. Input falls 12.5x, output 21.25x, and cached input 75x ^[4]. That last number should be read against the harness design: Muse Code runs subagents that persist for a session and keep what they have already learned about a repository instead of re-deriving it, according to Mark Zuckerberg ^[13]. The cheapest line on the contributor price list is exactly the re-served context that pattern generates ^[3]. The deepest cut lands on the tokens most likely to be somebody's proprietary code.

For a plain workload of a million input and a million output tokens with no caching, the standard tier costs $5.50 and the contributor tier $0.30, a factor of 18.3 ^[5]. Applied to the $0.40 per task Artificial Analysis measured at xhigh reasoning ^[8], that is about two cents a task ^[19], assuming the published figure reflects standard rates.

The cost case does not depend on the trade. On Vals AI's Finance Agent v2, Muse Spark 1.2 placed first of 45 models at 60.60 percent and $0.77 per task, against 58.63 percent and $5.12 for second-place Claude Opus 5, in roughly half the time per test ^[10]. That is already 6.6x cheaper than the model it beat ^[18]. Whichever tier those per-task numbers use, the gap is wide enough that handing over prompts is not what makes this affordable.

One input does not move: web search stays at $2.50 per thousand queries ^[6]. At contributor input rates that single line buys the equivalent of 25 million input tokens; at standard rates, 2 million ^[7]. Search-heavy agent runs will recover less of the discount than the token table implies.

Then there is the direction of disclosure. Meta has not published Muse Spark 1.2's parameter count, architecture, knowledge cutoff, training data or method ^[12], while posting a per-token price for yours. The one training input it did name is data from Muse Spark 1.1, its own previous model ^[22]. Zuckerberg has separately argued that US labs are disadvantaged by restrictions on training data ^[15], and high-quality code is the scarce part of the corpus ^[16]. The contributor tier reads as a way to buy that scarcity from the people who write it, priced in tenths of a cent.

The governance tooling that ships in the box is a side effect. Muse Code writes every model call, tool run, plan approval and file edit to a log on the user's machine, so it can resume from the step it reached after a crash ^[14]. That log is a recovery feature, and it is also the only local record of what was sent.

deeplearning.ai notes that OpenAI and others have tried similar initiatives, and that this discount could be transformative if it catches on ^[17]. What is new is the posted rate. Legal review now has a number to price against.