Most AI tools allow you to opt out of sharing your usage with model providers to improve future versions. Meta took that idea and put a price tag on it.
For new Muse Spark models for coding and other agent interactions, Meta offers an average of approximately 95% explicit discounts to users who “contribute” to the development of future models by sharing prompts and model outputs.
One million input tokens costs $1.25 under the standard contract, but only 10 cents under the contributor pricing model. The standard price for output tokens is $4.25 per million, while the same cost per million in the contributor model is only 20 cents.
Meta had a hard time getting training data. An initiative launched earlier this year to track employee computer usage drew widespread criticism within the company and was suspended in June. The company did not respond to TechCrunch’s questions about the new pricing model.
This type of user data is essential to improving the functionality of agent tools. “The reason we saw a big jump in[coding agent]functionality between April 2025 and October 2025 is because Claude Code saves all coding agent sessions by default and uses them for reinforcement learning training,” Mario Zechner, developer of the open source Harness Pi, told TechCrunch last month.
However, even though the adoption of agent tools for use outside of software engineering is increasingly essential for model builders, the ability to evaluate and improve these tools is hampered by the complexity of many professional workflows and the lack of digital traces.
Arvind Narayanan, a computer science professor at Princeton University, said there is ample evidence that large companies do not want their data used to train models.
“They’re sticking with their token-billed enterprise plans even though subscription-based consumer plans like Claude Max and ChatGPT Pro are getting discounts of 10x to 20x or more! (The main difference between both plans is data retention and enterprise IT governance),” he wrote on social media.
Perhaps recognizing this dynamic, Meta is offering companies explicit rewards for obtaining that information. Its pricing guide states that the contributor tier “lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on data is acceptable.”
Narayanan suggested this could incentivize large companies to look harder at what data is truly proprietary and what data can be shared with model providers.
This framework may also have an impact on increasing price competition among frontier labs. Anthropic’s latest Fable and Mythos models released yesterday had reduced processing costs for cached tokens, while OpenAI’s latest model received a significant price cut at the end of July.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
