Meta is offering a roughly 95% discount on its new Muse Spark model — built for coding and agentic workloads — to customers who let the company train on their prompts and outputs. Under the standard agreement, 1 million input tokens cost $1.25 and 1 million output tokens cost $4.25. Under the contributor tier, those same prices drop to 10 cents and 20 cents respectively. It is the most explicit price-for-data trade any frontier lab has put on a public rate card.
The pricing structure inverts the industry norm, where opting out of training is the default premium a customer pays for privacy. Meta has flipped that into a paid opt-in, with the discount sized to make the trade obvious. The company's own pricing guide frames the contributor tier as one that "lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable."
The move follows a bruising internal episode. Earlier in 2026, Meta launched an initiative to track how its own employees used computers, aiming to generate training data for agentic workflows. It drew wide internal criticism and was paused in June. Meta did not respond to a question about the new pricing model.
Key facts
- 01Meta's Muse Spark contributor tier charges 10 cents per million input tokens versus $1.25 under the standard agreement — a 95% discount.
- 02Output tokens drop from $4.25 per million to 20 cents per million for users who agree to share prompts and outputs for training.
- 03Meta's earlier employee-usage tracking initiative, launched in 2026, was paused in June after internal criticism.
- 04Anthropic's newly released Fable and Mythos models lowered cached-token pricing; OpenAI cut prices at the end of July.
Agentic training data has become the pinch point across the frontier. Model providers can benchmark chat and coding capabilities easily, but the digital exhaust from real professional workflows — multi-step tool calls, corrections, retries — is scarce and disproportionately valuable.
“The reason we saw a big jump in [coding agent] capabilities between April 2025 and October 2025 was that Claude Code, by default, would store all your coding agent sessions and use them for reinforcement learning training”— Mario Zechner, Developer behind open source harness Pi
That scarcity is exactly what powered Claude Code's rapid gains, according to Mario Zechner, the developer behind the open source harness Pi. Anthropic's coding agent stored user sessions by default and fed them into reinforcement learning, which Zechner credited for the jump in agent capability between April 2025 and October 2025. Muse Spark's contributor tier is Meta's attempt to buy the same flywheel rather than build it from employee monitoring.
The catch is that the customers with the most valuable data are precisely the ones least willing to share it. Arvind Narayanan, a Princeton computer science professor, pointed out that large enterprises routinely pay 10x to 20x more for token-billed enterprise plans rather than move to consumer subscriptions, because the main gap between the tiers is data retention and IT governance.
“They stick with token-billed Enterprise plans even though the subscription-based consumer plans like Claude Max and ChatGPT Pro are discounted by 10x-20x or even more! (The main difference between the plans is data retention + enterprise IT governance)”— Arvind Narayanan, Princeton computer science professor
Meta's bet is that a 95% price cut is large enough to move at least some of that data loose. For a startup running experiments where the prompts are not competitively sensitive, the math is straightforward: paying 10 cents instead of $1.25 per million input tokens changes what workloads are economical. Narayanan suggested the framework could push enterprises to more carefully sort which data is genuinely proprietary and which is fair game to trade.
It also lands in a market where the price of frontier inference is falling fast. Anthropic's newly released Fable and Mythos models cut costs for processing cached tokens. OpenAI pushed through significant price cuts on its latest models at the end of July. Muse Spark's headline discount now sits alongside those moves, but with a structural twist — the discount is conditional on the customer becoming a training-data supplier.
The risk for Meta is that the tier attracts the wrong workloads. If contributor-tier usage skews toward toy prototypes and synthetic tests, the reinforcement signal will be weak, and the discount becomes a straight subsidy without the training-data flywheel it was designed to fund. The company is also relying on customers to accurately self-sort, when in practice the boundary between "acceptable to train on" and "proprietary" is fuzzy inside most engineering organizations.
Meta has now put a public number on what agentic training data is worth to a frontier lab: roughly $1.15 per million input tokens and $4.05 per million output tokens in foregone revenue. That is a floor, not a ceiling, and it will drag every other lab's implicit pricing into the open. Expect Anthropic, OpenAI, and Google to be asked why their consumer tiers are already 10x to 20x cheaper than their enterprise tiers if the only real difference is data rights — and expect enterprise procurement teams to start negotiating on exactly that basis.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
