Most AI companies let you opt out of sharing your usage data. Meta is now paying you to opt in. For its new Muse Spark model, aimed at coding and other agentic workflows, Meta is offering a contributor pricing tier that cuts token costs by roughly 95% in exchange for allowing the company to use your prompts and model outputs to train future models. Input tokens drop from $1.25 per million to $0.10. Output tokens go from $4.25 per million to $0.20. That is not a modest discount. That is practically giving the API away.
As TechCrunch reported, Meta did not respond to questions about the new pricing structure, which makes the move feel more like a quiet experiment than a confident product decision. But the logic behind it is clear enough. Meta has been struggling to get quality training data, especially the kind of real-world interaction data that matters most for improving agentic systems.
Earlier this year, the company launched an internal initiative to monitor employee computer usage as a way to gather that data. It was paused in June after significant pushback from staff. The contributor pricing tier looks like the next attempt at solving the same problem, this time by going to paying customers instead of internal employees.
The data gap matters more than it might seem. Mario Zechner, developer of the open source agent harness Pi, told TechCrunch last month that a major jump in coding agent capabilities between April and October 2025 came largely because Claude Code was storing agent sessions by default and using them for reinforcement learning. That kind of feedback loop is hard to build without real usage data, and it compounds over time. Models that get more of it improve faster.
But agentic use cases outside of software engineering are harder to capture. Many professional workflows leave few digital traces, and the tasks are complex enough that automated evaluation is unreliable. Meta’s pricing approach is a direct attempt to buy access to exactly that kind of data from the developers and companies building on Muse Spark.
The catch is that most large enterprises do not want their data used for model training, and they are willing to pay a premium to avoid it. Princeton computer science professor Arvind Narayanan pointed out that big companies consistently choose token-billed enterprise plans over cheaper consumer subscriptions, even when the price difference is 10x to 20x or more. The main distinction between those plans is data retention and enterprise governance. Companies are already pricing in the cost of keeping their data private.
Meta’s contributor tier probably will not move those enterprise buyers. But it could attract startups, solo developers, and teams doing early-stage prototyping who care more about cost than data control. Meta’s own pricing guide frames it as lowering the barrier for “prototyping, testing integrations, and scaling experiments where training on your data is acceptable.” That framing is honest, and it suggests Meta knows exactly who this is for.
The broader context here is a pricing war that is accelerating across the frontier labs. Anthropic’s new Fable and Mythos models came with reduced costs for cached token processing. OpenAI cut prices significantly on its latest models at the end of July. Meta’s contributor discount is a more aggressive move, but it is structured differently because it exchanges price for something specific: access to usage data. That makes it less of a straight price cut and more of a data acquisition strategy with a discount attached. Whether developers bite will depend on how sensitive their workflows actually are, and how much that 95% discount is worth to them in practice.
