Meta discounts Muse Spark pricing for customers sharing AI data
Meta offers steep Muse Spark discount for contributor data
Meta is offering customers of its Muse Spark AI model substantially lower token prices if they agree to share prompts and model outputs for the development of future models. Muse Spark, intended to operate coding and other agents, carries an average discount of about 95% under the company’s contributor pricing tier.
Under Meta’s standard agreement, one million input tokens cost $1.25 and one million output tokens cost $4.25. In the contributor tier, those prices fall to $0.10 per million input tokens and $0.20 per million output tokens. Meta describes the option as a way to lower the barrier to prototyping, testing integrations and scaling experiments where training on customer data is acceptable.
Agent data is valuable, but enterprise controls remain central
The pricing makes explicit the exchange that is often implicit in AI services: a lower bill in return for data that can help improve future systems. This is particularly relevant for agentic tools, whose performance depends on more than isolated prompts and answers. Coding-agent sessions can provide traces of tool use, intermediate decisions and outcomes that help model builders evaluate and improve their products.
Mario Zechner, the developer of the open-source Pi harness, linked a major rise in coding-agent capabilities between April and October 2025 to Claude Code’s default collection of coding-agent sessions for reinforcement-learning training. However, the same kind of usable digital trace is harder to obtain for many professional workflows outside software engineering, limiting providers’ ability to assess and refine agents for those settings.
Meta has faced difficulty obtaining such material. An initiative launched earlier in the year to track employees’ computer use drew widespread internal criticism and was paused in June. The company did not respond to a question about the Muse Spark pricing model.
Price competition meets data governance
Princeton computer science professor Arvind Narayanan has noted evidence that large companies prefer token-billed enterprise plans to much cheaper consumer subscriptions because data retention and enterprise IT governance differ. In this market context, Meta’s Muse Spark agentic coding launch illustrates Meta’s push in agentic coding while the contributor tier creates a separate economic choice around data use.
The proposal also arrives amid competition on model costs. Anthropic’s newly released Fable and Mythos models introduced lower prices for processing cached tokens, and OpenAI cut prices for its latest models at the end of July.
For businesses, the practical implication is to classify agent workloads before pursuing the discount: reserve sensitive or proprietary workflows for arrangements with the required retention and governance controls, and consider contributor pricing only for data that the organization has determined can be used in model training.