Meta is done pretending data is free. The company’s new Muse Spark model comes with a contributor pricing tier that drops input costs from $1.25 to 10 cents per million tokens and output costs from $4.25 to 20 cents — roughly a 95% discount. The catch? You hand over your prompts and model outputs to train future versions.
This isn’t subtle. The pricing guide says the contributor tier “lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable.” That last clause is carrying a lot of weight. Meta has struggled to get this kind of data through other means. An internal program to track employee computer usage launched earlier this year drew enough criticism to get paused in June. So now the company is simply paying for access.
Why does this matter for agentic tools specifically? Mario Zechner, developer of the open source Pi harness, pointed to Claude Code’s default session storage as the reason coding agents jumped in capability between April and October 2025. Real usage data is what makes reinforcement learning work. The problem: most professional workflows outside software engineering leave no digital trace, and companies are reluctant to share what traces do exist. Princeton professor Arvind Narayanan notes that large firms stick with pricier Enterprise plans precisely because they don’t want their data retained for training — even when consumer plans are 10-20x cheaper.
Meta’s move reframes the entire data question as a pricing question. If a 95% discount is on the table, companies suddenly have a financial incentive to sort out what’s actually proprietary and what they’d happily trade for cheaper inference. It also lands amid a broader price war: Anthropic’s Fable and Mythos models just cut cached token costs, and OpenAI slashed prices in late July. The labs are competing on price and data simultaneously, and Meta just combined both levers into one offer.
