Meta is charging users less money if they agree to let the company use their prompts and model outputs to train future versions of its AI.
The company is applying this discount to its new Muse Spark model, designed for coding agents and similar tasks. Users who contribute data receive a price reduction averaging about 95% compared to standard rates.
Under normal terms, one million input tokens cost $1.25. Under the contributor pricing, that same volume costs 10 cents. Output tokens follow the same logic: the standard price is $4.25 per million, while contributors pay just 20 cents.
This approach comes after Meta faced internal resistance regarding how it gathers training material. An initiative earlier this year to monitor employee computer usage drew wide criticism and was paused in June. The company did not reply to a question from TechCrunch about this new pricing structure.
Data from real-world usage is essential for improving agentic tools. Mario Zechner, the developer behind the open source harness Pi, told TechCrunch last month that a significant jump in coding agent capabilities between April 2025 and October 2025 occurred because Claude Code stored all coding agent sessions by default and used them for reinforcement learning training.
However, the ability for builders to evaluate and improve these tools is often blocked by the complexity of professional workflows and the lack of digital traces outside software engineering. Arvind Narayanan, a computer science professor at Princeton, noted that large companies generally do not want their data used for model training.
“They stick with token-billed Enterprise plans even though the subscription-based consumer plans like Claude Max and ChatGPT Pro are discounted by 10x-20x or even more! (The main difference between the plans is data retention + enterprise IT governance),” he wrote on social media.
Perhaps recognising these dynamics, Meta is offering companies explicit compensation to obtain that information. Its pricing guide notes that the contributor tier lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on user data is acceptable.
Narayanan suggested this could incentivise large companies to be more diligent about distinguishing which data is truly proprietary and which can be shared with model providers.
The framework also plays into growing price competition between major labs. Anthropic released its newest Fable and Mythos models yesterday with lowered costs for processing cached tokens, while OpenAI’s latest models received major price cuts at the end of July.
What it means
Developers building agents now face a trade-off: pay full price for data privacy or accept a steep discount while allowing their work to train Meta’s models. This shifts the conversation from whether companies can afford to train AI to whether they are willing to share their proprietary workflows for a financial incentive.
