Meta priced your prompt data at $1.15 per million

Meta shipped Muse Spark 1.3 on September 2 with a second SKU sitting next to it: a Contributor tier at $0.10 per million input tokens and $0.20 per million output, against $1.25/$4.25 for the standard model. Same weights, same 1M-token context, roughly 21x cheaper on output. The difference is a data clause. Which means Meta has now published something most vendors keep implicit — a rate card for your prompt data.

What actually happened

Per OpenRouter's listing, Muse Spark 1.3 Contributor went live September 2 at $0.10 input / $0.20 output / $0.002 cache read per million tokens, with a 1,048,576-token context window and one condition stated plainly: "Prompts and outputs may be used to improve Meta's products." TechCrunch frames the tier as Meta lowering the barrier for "prototyping, testing integrations, and scaling experiments where training on your data is acceptable," and notes the company's earlier internal data-collection program was paused in June 2026 after employee pushback.

Do the subtraction and you get the number that matters. Meta is discounting $1.15 per million input tokens and $4.05 per million output tokens in exchange for the right to train on what passes through. That is the price it has set on your prompts and your model's answers. Not a privacy policy. A line item.

Why prompt data pricing matters for your business

Nothing here is scandalous. It is a trade, stated up front, and for a large class of work it's a good one. Prototyping, spike work, internal tooling, anything where the input is public documentation or synthetic test data — take the 21x. Refusing free money to protect prompts that contain nothing is not discipline, it's superstition.

The discipline is knowing which side of the line each workload sits on, before the cheap tier gets set as a default in a config file somewhere. Customer support transcripts, contract review, anything touching a client's pricing, roadmaps, patient or payment data — that stays on the paid tier, and someone should be able to say why in one sentence. The failure mode is not a dramatic leak. It's a junior dev switching a base URL to cut the bill, six months before anyone reads the terms.

Two things make this tractable. Route through a single interface you own, so the tier is a config value you can audit in one place rather than a decision scattered across a dozen call sites. And write down the classification — public / internal / client-confidential — next to each workflow that calls a model. It takes an hour and it turns "are we training someone's model on our client data?" from an investigation into a lookup.

Meta was unusually honest about the exchange rate. Most vendors bury the same trade in section 4 of the terms.

Sources: OpenRouter, TechCrunch.