Meta Offers Steep Discounts to Users Who Share Data With Its New AI Model

Meta has rolled out a new pricing structure for its Muse Spark model, built to power coding and other AI agents. The company is offering a steep discount — averaging around 95% — to customers who agree to "contribute" to future model development by allowing their prompts and outputs to be used for training.

Under the standard agreement, one million input tokens cost $1.25, but under the contributor pricing tier, the same amount costs just $0.10. For output tokens, the standard rate is $4.25 per million, compared to only $0.20 under the contributor plan.

A Struggle for Training Data

Meta has previously faced difficulties collecting training data. An initiative launched earlier this year to monitor employees' computer usage drew significant internal backlash and was paused in June. The company did not respond to a request for comment on the new pricing model.

According to Mario Zechner, developer of the open-source harness Pi, a major leap in coding agent capabilities between April and October 2025 came because Claude Code, by default, stored all user coding sessions and used them for reinforcement learning training.

Princeton computer science professor Arvind Narayanan has pointed out that large companies often choose more expensive, token-billed enterprise plans over cheaper consumer subscriptions specifically to avoid having their data used for training. He suggested Meta's approach could push companies to more carefully assess which data is genuinely sensitive versus shareable.

The move comes amid broader price competition among leading AI labs — Anthropic recently lowered costs for cached tokens on its newest models, while OpenAI cut prices significantly on its latest models in late July.