Meta is paying to peek at how you use their latest AI model

Tim Fernholz reports that Meta’s new Muse Spark model — aimed at coding and other agents — sells a "contributor" tier that discounts usage for customers who share prompts and model outputs to help build future models. Average discount is about 95%: 1 million input tokens drop from $1.25 to $0.10; 1 million output tokens drop from $4.25 to $0.20.

Most AI tools let users opt out of training use. Meta inverted the frame and put a price tag on contributing. The company has struggled to get training data; an earlier push to track employees’ computer usage drew wide internal criticism and was paused in June. Meta did not answer TechCrunch’s question about the new pricing. Princeton’s Arvind Narayanan notes enterprises often stay on expensive token-billed plans rather than cheap consumer subscriptions precisely to avoid data retention — a 10x–20x gap that Meta is now trying to buy past with an explicit contributor discount.

Signal Desk’s read: this is Meta pricing the thing every lab wants and few will admit is scarce — high-quality agent traces. Mario Zechner’s point in the piece, that Claude Code’s default session storage fed a capability jump via RL between April and October 2025, is the competitive context. Agentic improvement increasingly depends on real workflow residue, not static benchmarks. Meta is not inventing that dependency; it is publishing a tariff for it.

The contributor guide’s own language — lower barrier for prototyping and experiments "where training on your data is acceptable" — is the product honesty. Enterprises that treat prompts as proprietary will stay on standard rates. Teams shipping greenfield agents may take the 95% cut and decide later what "acceptable" meant. Narayanan’s suggestion that explicit compensation could push large companies to separate truly proprietary data from shareable data is the optimistic read. The pessimistic read is that discount pressure slowly normalizes training on customer traffic the way "improve the product" toggles once did, only louder.

Who gains: Meta’s model quality loop if enough volume opts in; startups and hackers for whom $0.10 / $0.20 per million tokens changes experimentation math. Who should pause: any org whose legal team has not mapped which prompts may leave the building. Meta’s paused employee-monitoring episode is the cultural backdrop — desperation for interaction data, now redirected at paying customers with a carrot instead of an internal stick.

Price competition among frontier labs (Anthropic’s Fable/Mythos cached-token cuts; OpenAI’s late-July cuts) makes the contributor tier look like Meta fighting on two axes at once: sticker price and data acquisition. That combination is more aggressive than a pure list-price war.

Agent builders have been arguing that session logs are the scarce training resource for coding agents. Consumer plans already trade lower price for retention; Meta’s contributor tier makes the same trade explicit for API-style Muse Spark usage.