Meta Put a Price on Your Code: 12x Off If Meta Can Train on It

Meta shipped two things on August 5: Muse Code, a terminal coding agent, and Muse Spark 1.2, the model that drives it. The model is Meta's third release in four months. The agent is competent. Neither is the interesting part.

The interesting part is that the price list has two columns.

Standard tier: $1.25 per 1M input tokens, $0.15 on a cache hit, $4.25 output, 1M-token context, a rate ceiling around 3,000 requests per minute.

Contributor tier: $0.10 input, $0.002 cached, $0.20 output — 12.5x cheaper on input, 21.25x cheaper on output, capped at 60 requests per minute. Same weights, same context window, same model. The only difference is that Meta may train future models on your traffic.

Every major lab has run some version of this trade. It normally lives in a consumer terms-of-service, or in the fine print of a free tier, or in an enterprise contract's opt-out checkbox — always framed as a privacy setting, never as a line item. Meta moved it onto the rate card and attached a number to it.

That number is now the market's first explicit quote for what a million tokens of your engineering work is worth to a frontier lab: about $1.15 on input, $4.05 on output.

The rate cap is the real specification

A 60-requests-per-minute ceiling is the tell.

If Meta wanted volume, it would not cap the cheap tier at a rate that makes production fan-out impossible. Sixty rpm cannot serve a fleet of parallel agents. It can comfortably serve one developer, or one small team, running one coding agent against one real repository.

Which is exactly the data that is hard to buy.

Public code is exhausted — GitHub has been scraped to the studs, and the marginal value of another million lines of open-source Python is near zero. What labs cannot scrape is the process: the failed first attempt, the stack trace pasted back in, the correction, the tool call that returned the wrong file, the second correction, the passing test. Agentic traces on private codebases, produced by a human who cared whether the output worked.

Synthetic data does not have that shape. Neither does a benchmark. The contributor tier's rate limit is not a throttle on cost — it is a filter that selects for supervised, human-in-the-loop, real-repository traces and screens out bulk automated traffic that would dilute the corpus.

Meta priced the discount so it is irresistible to the exact population whose data it wants, and useless to everyone else.

What you get for the two cents

Muse Spark 1.2 scores 54 on Artificial Analysis's Intelligence Index at its highest reasoning setting. That is up from 51 for Muse Spark 1.1 and 43 for Muse Spark 1.0 in April — eleven points in four months.

It lands tied with Grok 4.5 (54), a point behind GPT-5.5 at xhigh (55), and behind the current top of the board: Claude Opus 5 (61), Claude Fable 5 (60), GPT-5.6 Sol (59), and Kimi K3 (57). On GDPval-AA v2 it holds an Elo of 1631, fifth overall.

On coding specifically — the axis Meta scaled training compute against — the numbers are close but not identical depending on who ran them. Meta reports 82.9% on Terminal-Bench 2.1 and 59.3% on DeepSWE v1.1. Artificial Analysis's independent run puts Terminal-Bench v2.1 at 80%. Elsewhere: SciCode 56%, Humanity's Last Exam 44%, τ³-Banking 27%, AA-Omniscience 22, CritPt 18%.

That is a genuinely strong mid-frontier model. It is not the best coding model available, and Meta is not claiming it is.

Muse Code, the agent, was co-trained with the model rather than layered on afterward. It runs persistent asynchronous background subagents and ships three bundled commands: /plan for approval-gated planning, /grill for stress-testing a change, /goal for run-to-completion. Meta's own case study walks it through GPU kernel optimization across more than 1,000 tool calls against a baseline implementation.

A thousand tool calls is a lot of trace.

The arithmetic that makes this work

Take an ordinary agentic turn: 50,000 tokens of repository context and system prompt in, 2,000 tokens out.

On the standard tier that is roughly $0.07. On the contributor tier, roughly $0.005.

Multiply by a few hundred turns a day across a team and the gap stops being a rounding error and starts being the line in the budget that someone has to defend. And the comparison that matters is not Meta-versus-Meta. DeepSeek's V4-Flash-0731 blends to about $0.06 per 1M tokens, and Chinese open weights are the thing currently setting the floor on agentic coding.

Meta cannot reach that floor on efficiency. Standard-tier Muse Spark 1.2 is twenty times above it. So Meta reached it a different way: it subsidized the gap and billed the difference in training data.

That is a strategically coherent move for the one frontier lab that does not sell inference as its primary business. OpenAI and Anthropic need API margin. Meta needs the next model to be better than this one, and it is willing to run the coding API at or below cost to get the corpus that makes that true.

What you are actually agreeing to

Read the trade plainly. In exchange for a 12.5x discount, Meta may train on your prompts, your completions, and the traces in between — which, for a terminal agent operating inside a repository, means your source files, your architecture, your internal conventions, your bug patterns, and the sequence of decisions your engineers made while fixing them.

For a side project, that is a fine deal and an obvious yes. For a company whose codebase is the product, it is a licensing decision wearing a pricing decision's clothes, and it will get made by whoever holds the API key rather than by anyone who would normally sign off on it.

Meta deserves credit for the clarity. It is easier to evaluate an explicit price than a buried consent. But clarity does not change what the transaction is.

The discount is not a discount. It is a purchase order, and you are the vendor.