Meta ties Muse Spark's cheapest tier to handing over your prompts
Meta's new contributor tier for its Muse Spark 1.3 coding model cuts token prices by roughly 95 per cent, and cached input tokens by about 75 times, in exchange for letting Meta train on developers' prompts and outputs. Meta calls the tier a way to lower the barrier to prototyping and testing; a Princeton researcher says large companies will likely stick with costlier enterprise plans that keep their data out of training, for governance reasons.
Artificial Intelligence··Morning
Contributor pricing trades a discount for training data
Meta has added a tier it calls contributor pricing to Muse Spark, the model it built for coding and agent applications. On the standard tier, 1 million input tokens cost 1.25 dollars and 1 million output tokens cost 4.25 dollars; on the contributor tier the same amounts cost 0.10 dollars and 0.20 dollars, a discount of roughly 95 per cent. The prompts and outputs collected under the contributor tier are used in reinforcement learning to improve later versions, and Meta's own pricing guide describes the tier as a way to lower the barrier to prototyping, integration testing and scaling experiments where training on a developer's data is acceptable.[1]
Cached tokens fall by 75 times under the same trade
WeRSM's read of Meta's published price sheet adds a category the standard report does not cover: 1 million cached input tokens cost 0.15 dollars on the standard tier and 0.002 dollars under the contributor tier, a cut of roughly 75 times. The outlet calls the arrangement one of the clearest cases yet of an AI company pricing access to user data directly, and notes that Muse Spark 1.3 is aimed at coding and agentic workflows, where an agent inspecting files, calling tools and correcting errors produces a long interaction log that a benchmark alone cannot supply.[2]
Big customers are expected to skip the trade anyway
Both accounts describe the same trade from different sides of the price sheet: standard-tier traffic is excluded from training while contributor-tier prompts and outputs help build later models, and the contributor discount runs from roughly 95 per cent on ordinary tokens to about 75 times on cached ones. Princeton researcher Arvind Narayanan, cited in the standard report, expects large companies to skip the contributor tier anyway: he says such companies tend to pick the costlier enterprise plans instead, citing data retention and enterprise IT governance.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
