Meta Offers a 95 Percent Discount on Its Muse Spark AI Model for Developer Data

Meta is offering developers of its Muse Spark coding model steep discounts, an average of about 95 percent by TechCrunch’s estimate, if they agree to let the company train future models on their prompts and outputs, according to TechCrunch. The arrangement inverts the usual privacy choice most AI providers offer: instead of letting customers opt out of data collection for free, Meta is charging a premium for privacy and discounting the alternative.

The pricing applies to Muse Spark 1.3, Meta’s newest coding and agentic model, which began rolling out through the Muse Code terminal agent and Meta’s Model API on September 2, according to Meta’s own research blog. Under Meta’s standard agreement, which does not use a customer’s prompts or outputs for training, TechCrunch reports Meta prices access at $1.25 per million input tokens and $4.25 per million output tokens. Under the “contributor” tier, which grants Meta permission to train on that traffic, the same million input tokens cost just 10 cents and the same million output tokens cost just 20 cents: discounts of roughly 92 percent and 95 percent.

What the Discount Actually Buys Meta

TechCrunch reports that Meta’s own pricing guide describes 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.” In practice, that means a developer’s prompts and the model’s responses to them, not the model’s weights or Meta’s own source code, become training material for whatever comes after Muse Spark 1.3.

The tier comes with real constraints beyond the training trade-off. Meta’s developer documentation describes the contributor tier as “rate-limited by tokens in a rolling 5-hour window, not by request count,” a different throttling model than the standard tier uses, and states that contributor access is “available in select countries.” Several third-party pricing trackers have also reported a fixed requests-per-minute ceiling on the contributor tier, but that detail does not appear in Meta’s own documentation, which describes the throttle only in token terms.

A Second Attempt at Getting the Data

Meta has had a rough time trying to obtain training data by other means. TechCrunch notes that an internal initiative to track the computer usage of Meta’s own employees, launched earlier this year, “attracted wide internal criticism and was paused in June.” The contributor tier pursues a similar goal through outside developers instead, replacing an internal mandate with an explicit, opt-in price.

The underlying scarcity is real. Mario Zechner, the developer behind the open source coding harness Pi, told TechCrunch last month that “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.” Meta’s contributor tier is an attempt to buy that same kind of real-world usage data on purpose, rather than collect it by default the way Zechner described.

Why Big Customers May Still Say No

Arvind Narayanan, a Princeton computer science professor, argued in a social media post quoted by TechCrunch that large companies already show they will pay to keep their data out of a model provider’s training pipeline, even against a steep discount. Enterprise customers, he wrote, “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).” Narayanan suggested pricing frameworks like Meta’s contributor tier could still matter at the margins, by pushing large companies to be more diligent about which of their data is truly proprietary and which could safely be shared with a model provider.

Part of a Wider Price Race

Meta’s contributor tier lands in the middle of an active round of price cuts across the major AI labs. TechCrunch notes that Anthropic’s newest Fable and Mythos models had just shipped with lower prices for processing cached tokens, and that OpenAI made major price cuts to its own latest models at the end of July. Meta did not respond to TechCrunch’s question about the new pricing model.