Meta is now offering customers of its new Muse Spark AI model a discount averaging roughly 95% in exchange for permission to use their prompts and outputs to train future versions, according to a report published Wednesday by TechCrunch. The Muse Spark model is designed for running coding and other autonomous agents. While most AI platforms let users decline to share their usage data with the provider for model improvement, Meta has attached an explicit financial incentive to participation.
Under the contributor pricing structure, one million input tokens cost just 10 cents, compared to $1.25 under Meta's standard agreement. Output tokens are priced at 20 cents per million tokens for contributors, versus $4.25 per million in the standard tier. The contributor tier is positioned to reduce obstacles for prototyping, testing integrations, and expanding experiments in situations where training on customer data is acceptable, the pricing guide states.
The company has struggled to secure training data through other means, according to the report. An effort launched earlier this year to monitor the computer activity of Meta's own employees drew broad internal pushback and was suspended in June. Meta did not reply to TechCrunch's inquiry about the new pricing approach. Mario Zechner, developer behind the open source harness Pi, told TechCrunch last month that user data of this kind is essential for improving agentic tools, noting that "the reason we saw a big jump in [coding agent] capabilities between April 2025 and October 2025 was that Claude Code, by default, would store all your coding agent sessions and use them for reinforcement learning training."
This user information becomes increasingly critical as model builders shift toward deploying agentic tools beyond software engineering, the report explains, yet their capacity to assess and enhance those tools is hampered by the intricate nature and absence of digital records for many professional workflows. Princeton computer science professor Arvind Narayanan observed that strong evidence suggests large companies resist having their data used for model training, continuing to pay for token-billed enterprise plans even though subscription-based consumer plans like Claude Max and ChatGPT Pro offer discounts of 10 to 20 times or more. The primary distinction between these plans is data retention plus enterprise IT governance. Meta's explicit compensation framework could encourage major enterprises to be more careful about distinguishing truly proprietary data from information that could be shared with model providers. The structure may also feed into intensifying price competition among leading AI labs, as Anthropic's newest Fable and Mythos models released yesterday included reduced costs for processing cached tokens, while OpenAI's latest models received significant price cuts at the end of July.
By making data contribution economically attractive rather than merely optional, Meta is testing whether corporate customers will trade proprietary information for substantial cost savings in a market where training data has become a bottleneck for the next generation of AI tools. The framework signals that access to real-world usage patterns may become as much a currency as cash in enterprise AI contracts. If this approach gains traction, procurement teams may soon face a new calculus where the value of their operational data becomes a negotiating lever rather than simply a privacy concern.