Meta Cuts Muse Spark Price in Exchange for Data Sharing

Meta associates Muse Spark with a "contributor" rate that drops by about 95% if the user agrees to share their queries and outputs for future training. This proposal comes at a time when large accounts are avoiding having their data used for training, and price reductions are multiplying at Anthropic and OpenAI.

Cautious Companies and Price Wars

Arvind Narayanan, a computer science professor at Princeton, believes there is solid evidence that large companies prefer to avoid having their data used to train models. He observes that these companies maintain enterprise offerings billed per token, even as consumer subscriptions like Claude Max and ChatGPT Pro see their prices drop by a factor of 10 to 20 or more, with the major difference concerning data management and IT governance. Narayanan suggests that the compensation offered by Meta could encourage these companies to better distinguish proprietary data from data that can be shared with model providers. Meanwhile, Anthropic has released the Fable and Mythos models with reduced costs for processing cached tokens, and OpenAI's latest models have seen significant price drops at the end of July.

Muse Spark: Massive Discounts for Contribution

Meta applies an explicitly reduced rate for Muse Spark users who agree to share their queries and model outputs to contribute to the training of future versions. Muse Spark is a new model from Meta designed in part to operate coding agents and other agents. The discount reaches about 95%: one million input tokens costs $1.25 at standard rates, but only 10 cents for contributors; for output tokens, the price drops from $4.25 to 20 cents per million. Meta's pricing guide specifies that this "contributor" level aims to facilitate prototyping, integration testing, and the expansion of experiments when training on user data is acceptable. Most AI tools also allow users to refuse sharing usage data with the model provider.

Why Usage Data is in Demand

Usage data is essential for improving the performance of agent tools. Mario Zechner, who develops the open-source tool Pi, notes that between April 2025 and October 2025, the performance of the coding agent significantly increased because Claude Code automatically recorded all coding agent sessions and used them for reinforcement training. However, the development of these tools in areas other than software engineering remains hindered by complexity and the lack of digital traces in many professional processes.

An Internal Precedent and Unanswered Questions

Meta sought to collect more training data by implementing, earlier this year, a tracking system for activity on its employees' computers. This initiative sparked significant internal backlash and was halted in June. When asked about its new pricing based on data contribution, the company did not respond.