Meta Offers 92-95% Discount on Muse Spark AI Model to Users Who Share Their Prompts

Meta is charging customers a steep premium to keep their prompts and outputs private when using its new Muse Spark model, built for operating coding and other AI agents.

Under Meta's published pricing guide, standard usage of Muse Spark costs $1.25 per million input tokens and $4.25 per million output tokens. But customers who agree to let Meta use their prompts and model outputs to train future models — a "contributor" tier — pay just $0.10 per million input tokens and $0.20 per million output tokens. That's a roughly 92% discount on input and a 95% discount on output.

Meta did not respond to a request for comment on the pricing structure.

The scheme is Meta's latest attempt to solve a persistent data problem. Earlier this year, the company launched an initiative to track employees' computer usage as a source of training data, but paused it in June after internal criticism.

Usage data has become central to improving agentic AI tools. Mario Zechner, developer of the open-source coding harness Pi, told TechCrunch last month that a jump in coding-agent capability between April and October 2025 came largely because Claude Code stored user sessions by default and used them for reinforcement learning. Model builders increasingly need this kind of data to improve agents for use beyond software engineering, but many professional workflows lack the digital traces needed to train and evaluate on.

Large enterprise customers, however, have shown a strong preference for keeping their data out of training pipelines. Princeton computer science professor Arvind Narayanan noted on social media that big companies routinely stick with token-billed enterprise plans rather than switching to consumer subscription plans like Claude Max or ChatGPT Pro, even though those plans can be discounted 10x to 20x or more. The primary distinction between the plans, he wrote, is data retention and enterprise IT governance — not raw capability.

Meta's contributor pricing appears designed to make that trade-off explicit, offering direct compensation rather than leaving data-sharing as a passive default. Meta's own pricing guide describes the contributor tier as a way to "lower the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable." Narayanan suggested the framework could push large companies to more carefully sort which data is genuinely proprietary and which could safely be shared.

The pricing move lands amid broader cost competition among frontier labs. Anthropic's newly released Fable and Mythos models introduced lower prices for cached tokens, while OpenAI cut prices significantly on its latest models at the end of July.

What this means

Meta's contributor tier turns an opt-out checkbox into a real financial decision, effectively pricing in the value of user data rather than asking for it for free. It's a tacit admission that internal data-collection efforts, like the paused employee-monitoring program, haven't solved Meta's training-data gap for agentic use cases. Whether enterprises take the discount will depend on how much they trust Meta's data-handling commitments — and how much they value keeping their proprietary workflows out of a competitor's training set.