Meta cuts Muse Spark token prices by about 95% for users who share their data

A discount in exchange for your data

Meta has attached a price tag to a decision most AI providers leave as a simple toggle. For its newly launched Muse Spark model, built for coding and other agentic workloads, the company is offering customers who let it collect their prompts and model outputs for future training a discount that works out to roughly 95% off standard rates, according to TechCrunch.

The scheme is unusual because of how explicit it is. Most AI tools let users opt out of sharing usage data with the provider; Meta has instead turned that choice into a commercial transaction, paying for data in the form of lower token prices.

What the two tiers cost

Under the standard agreement, TechCrunch reports, one million input tokens costs $1.25 and one million output tokens costs $4.25. Under the contributor tier, the same volumes cost 10 cents and 20 cents respectively. Meta's pricing documentation pitches the cheaper tier as a way to lower the barrier to prototyping, testing integrations and scaling experiments in situations where training on a customer's data is acceptable. The company did not respond to a question from TechCrunch about the new pricing model.

Why agents need real usage traces

The pricing reflects how valuable real-world agent data has become. According to Mario Zechner, the developer behind the open-source harness Pi, the noticeable jump in coding-agent capabilities between April 2025 and October 2025 came largely because Claude Code stored users' coding sessions by default and used them for reinforcement learning training.

Usage data of this kind is harder to come by outside software engineering, TechCrunch notes, because many professional workflows are complex and leave few digital traces, which limits how well model builders can evaluate and improve agentic tools for general use. Meta has also struggled internally: an initiative launched earlier this year to track its own employees' computer usage drew wide internal criticism and was paused in June.

The enterprise holdout

Arvind Narayanan, a computer science professor at Princeton, pointed out that there is good evidence large companies do not want their data used for model training. He observed that such firms stick with token-billed enterprise plans even though subscription-based consumer offerings like Claude Max and ChatGPT Pro are discounted by a factor of ten to twenty or more, with the main differences being data retention and enterprise IT governance.

By offering explicit compensation, Meta may be acknowledging that dynamic rather than fighting it. Narayanan suggested the arrangement could actually encourage large companies to be more careful about distinguishing which data is genuinely proprietary and which could be shared with model providers in return for lower costs.

Part of a broader price war

The move also lands amid intensifying price competition among frontier labs. As TechCrunch reports, Anthropic's newest Fable and Mythos models, released the day before the article, came with reduced costs for processing cached tokens, while OpenAI delivered major price cuts on its latest models at the end of July. Meta's contributor tier adds a new dimension to that competition: instead of simply cutting list prices, it is effectively subsidising them with data.

Why it matters

The scheme converts a privacy setting into a measurable financial quantity. If a customer's prompts and outputs are worth around 95% of their token bill, both sides now have to treat usage data as an asset with a going rate rather than fine print. That could reshape how startups and individual developers budget for frontier models, since the discount is large enough to change which provider they can afford. It may also push enterprises toward a more honest accounting of what in their workflows is actually sensitive. The open question is whether other labs follow suit, and whether users who take the discount fully grasp that their agent sessions, including potentially proprietary work, become raw material for the next generation of models.