Meta pays users to share AI model usage data under new Muse Spark pricing

TechCrunch reports that Meta has introduced a pricing scheme for its new Muse Spark model, which is intended for operating coding and other agents, that gives users a discount in return for sharing their data. Most AI tools let customers opt out of having their usage used to improve future models; Meta has attached a price to that choice. Users who agree to contribute their prompts and model outputs to the development of later models receive a discount averaging about 95%.

The article gives the figures. Under the standard agreement, one million input tokens cost $1.25, while under the contributor model they cost 10 cents. One million output tokens cost $4.25 at the standard rate and 20 cents as a contributor. Meta's pricing guide says the contributor tier makes it cheaper to prototype, test integrations and expand experiments when the customer accepts training on its data. TechCrunch notes that Meta did not answer its question about the model. It also says Meta has struggled to obtain training data: an effort to track employees' computer use was widely criticised inside the company and paused in June.

TechCrunch explains why such data matters. Developer Mario Zechner is quoted attributing the leap in coding agent ability from April to October 2025 to Claude Code keeping sessions by default for reinforcement learning. Evaluating agents in other professional workflows is harder because few digital traces exist. Princeton professor Arvind Narayanan observes that large companies tend to keep token-billed enterprise plans, whose main difference is data retention and IT governance, and suggests Meta's offer could push firms to be more careful about which data is truly proprietary. The article adds that the scheme comes amid price competition among frontier labs.