Meta is Funding Access to Analyze Your Usage of Their Latest AI Model

Many AI tools give users the option to prevent sharing their usage data with the model providers to enhance future iterations. Meta has taken this concept and added a monetary incentive.

For its new Muse Spark model, designed to run coding and other types of agents, Meta is offering substantial discounts averaging around 95% for users who choose to “contribute” by sharing their prompts and model outputs to aid in developing future models.

Under a standard agreement, 1 million input tokens are priced at $1.25, but through the contributor pricing model, that cost drops to only 10 cents. For output tokens, the typical charge is $4.25 per million; under the contributor model, that same million costs just 20 cents.

Meta has faced challenges in acquiring training data: An effort to monitor its employees’ computer usage, initiated earlier this year, faced significant internal backlash and was halted in June. The company did not respond to TechCrunch’s inquiry regarding its new pricing strategy.

Such user data is crucial for enhancing the performance of agentic tools. “The significant improvement we observed in [coding agent] capabilities between April 2025 and October 2025 was due to Claude Code defaulting to store all your coding agent sessions for reinforcement learning training,” explained Mario Zechner, the developer of the open source harness Pi, in a conversation with TechCrunch last month.

However, as the demand for model creators to deploy agentic tools beyond software engineering grows, their capacity to assess and improve these tools is hampered by the complexities and the absence of digital traces in various professional workflows.

Arvind Narayanan, a professor of computer science at Princeton, pointed out that there’s strong evidence suggesting large companies prefer to keep their data out of model training processes.

“They continue with token-billed Enterprise plans despite subscription-based consumer options like Claude Max and ChatGPT Pro being discounted by 10x-20x or more! (The primary distinction between the plans is data retention and enterprise IT governance),” he noted on social media.

In light of these dynamics, Meta is incentivizing companies with explicit compensation to obtain that data. Its pricing guide states that the contributor tier “reduces the barrier to entry for prototyping, testing integrations, and scaling experiments where utilizing your data for training is permissible.”

Narayanan suggested that this could motivate large companies to take more care in identifying which data is genuinely proprietary and which could potentially be shared with model providers.

This framework might also contribute to escalating price competition among frontier labs. Anthropic’s latest Fable and Mythos models, released yesterday, offered reduced costs for processing cached tokens, while OpenAI’s recent models received significant price reductions at the end of July.