AINEWS
Meta's 95% discount on its new Muse Spark tier is paid for with prompts that train future models, inverting the enterprise deal where companies pay extra to keep their data out of training.
Meta's new Muse Spark "contributor tier" cuts token prices by about 95%. Standard input on the model costs $1.25 per million tokens; the contributor tier charges 10 cents. Standard output runs $4.25 per million; the contributor tier charges 20 cents.
The trade sits in the pricing page: anyone who sends a prompt through the contributor tier agrees that Meta can use both the prompt and the model's response to train future versions of the model. There is no per-conversation opt-out, no flag for proprietary code, no list of categories Meta will exclude. The discount is the price of unconditional training-data consent.
Meta's pricing guide, per TechCrunch, frames the contributor tier as a way to "lower the barrier to entry for prototyping, testing, and scaling where training on user data is acceptable." The phrase "where training on user data is acceptable" is the operative one. It signals who Meta expects to use the tier: developers and tinkerers running workloads they would not mind seeing again in a future model.
Enterprises pay the list price for AI access because they do not want their prompts used for training. Arvind Narayanan, a Princeton computer science professor, has argued that large companies prefer token-billed Enterprise plans over discounted consumer subscriptions largely because of data retention, per TechCrunch. Meta's contributor tier is the inverse: a steep discount in exchange for surrendering the data that enterprises pay extra to keep.
The new pricing structure lands weeks after a quieter retreat inside Meta. In June 2026, the company paused an internal initiative that would have tracked employees' computer usage, after internal criticism and unresolved data-security concerns, Reuters reported. The publicly stated goal of the program was to source training data from inside the company. Its collapse left a hole in the data pipeline that public-user data is now being asked to fill.
Other frontier labs already use coding-agent sessions for reinforcement-learning training, but none have published a list price for the trade. Mario Zechner, the developer of the open-source Pi coding harness, has attributed the leap in coding-agent capabilities between April and October 2025 to a default behavior in Claude Code: every coding-agent session was, by default, stored for training, TechCrunch reported. The data was the lever, and the contributor tier makes the lever visible to anyone reading the pricing page.
The deal also shows where its own limits sit. Contributor-tier traffic has no training opt-out, so any prompt a developer sends is, by default, training data. Code that includes customer secrets, internal APIs, regulated data, or anything covered by an NDA does not belong in the tier. Meta did not respond to TechCrunch's questions about whether it will exclude any category of input, and the pricing guide lists none.
The companies most likely to take the discount are the ones for whom the training-data trade is genuinely cheap. That is a small set: teams with little proprietary code, workloads whose prompts would not be missed in a future training set, and use cases where a future model's familiarity with the same prompts is a feature rather than a leak. It is also the set whose data is least likely to move the next model forward. Meta is, in effect, paying full price for the data it cannot get from employees, and 5% of list for the data it can get from everyone else.
A portable framework, then, is short: in AI's data-hunger economy, the size of a discount is the size of the data the lab wants from you. A 95% discount is not a price war. It is a price.