While many AI services offer an option to prevent your data from being used for model improvement, Meta is now monetizing this choice.
Specifically for its new Muse Spark model, designed for coding and agent operations, Meta is providing a significant discount, roughly 95%, to users who agree to share their prompts and model outputs, thereby “contributing” to the model’s future development.
Under the standard pricing structure, a million input tokens cost $1.25, but with the contributor model, the price drops to just $.10. Similarly, output tokens, normally priced at $4.25 per million, are available for only $.20 per million for contributors.
Meta has faced challenges in acquiring training data; for instance, an earlier plan to monitor employee computer activity this year drew significant internal backlash and was halted in June. TechCrunch’s inquiry regarding this new pricing strategy received no response from the company.
Such user data is crucial for enhancing the performance of agentic tools. Mario Zechner, creator of the open-source harness Pi, informed TechCrunch last month that “the significant improvement in [coding agent] capabilities observed between April and October 2025 stemmed from Claude Code’s default practice of retaining all coding agent sessions for reinforcement learning training.”
However, despite the growing need for model developers to implement agentic tools beyond software engineering, the intricate nature and scarcity of digital records in numerous professional fields hinder their capacity to assess and refine these tools.
Princeton computer science professor Arvind Narayanan highlighted substantial evidence indicating that major corporations prefer their data not be utilized for model training.
He stated on social media that “These companies opt for token-billed Enterprise plans, even though consumer-focused subscription plans such as Claude Max and ChatGPT Pro offer discounts of 10x to 20x or greater! (The primary distinction between these plans involves data retention and enterprise IT governance).”
Potentially acknowledging these market behaviors, Meta is now directly compensating companies for this data. According to its pricing guide, the contributor tier “simplifies prototyping, integration testing, and scaling experiments in scenarios where using your data for training is permissible.”
Narayanan proposed that this approach might encourage large companies to more carefully distinguish between truly proprietary data and data that could be safely shared with model providers.
This pricing model could also intensify the ongoing price competition among leading AI research labs. For instance, Anthropic’s recently launched Fable and Mythos models introduced reduced costs for cached token processing, and OpenAI’s newest models saw significant price reductions in late July.