Meta cuts Muse Spark prices 92 percent for users who allow training

TechCrunch reports that Meta has launched an explicit contributor pricing tier for its new Muse Spark model that offers roughly a 92 to 95 percent discount to developers who let the company use their prompts and outputs to train future models. The vendor's own pricing documentation confirms that a standard tier and a contributor tier now exist side by side on the Meta Model API.

The numbers TechCrunch reports

Under the standard agreement, one million input tokens cost $1.25 and one million output tokens cost $4.25. Under the contributor pricing model, the same million input tokens cost 10 cents and the same million output tokens cost 20 cents. That is a 92 percent discount on input and a 95 percent discount on output. Meta did not respond to a TechCrunch question about the pricing model.

What the vendor pricing page says

Meta's own developer pricing page describes itself as covering "standard and contributor pricing tiers, per-token pricing, image pricing, Muse Voice Transcribe pricing, and rate limits for Meta Model API." The contributor tier is described in Meta's guidance as one that "lowers the barrier to entry for prototyping, testing integrations, and scaling experiments where training on your data is acceptable." Meta has not published a separate independent breakdown of what data is retained, or for how long, beyond that tier description.

Why Meta wants the data

TechCrunch quotes Mario Zechner, the developer behind the open-source coding harness Pi, on why session data has become the key input to improving agent models. Zechner told the outlet last month that the sharp rise in coding agent capabilities between April and October 2025 was because Claude Code stored coding sessions by default and used them for reinforcement learning training. Meta itself has clearly wanted training material. An internal initiative to track the computer usage of Meta employees, launched earlier this year, was paused in June after wide criticism.

Enterprises have been paying the premium

Princeton computer science professor Arvind Narayanan told TechCrunch that large companies already show their choice through what they pay. He noted that enterprises "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," and said the main difference between those plans is data retention and enterprise IT governance. In that reading, Meta's tier structure puts a public number on the premium enterprises have been quietly paying.

The interesting change is separating the price of the tokens from the price of the data, and making both visible. A team on the standard tier is now paying an explicit twelve times the input rate to keep its prompts. That decision has to be made for each agent individually, and the honest answer depends on which prompts carry proprietary reasoning, which carry customer data, and which are cheap boilerplate. The other consequence is competitive. Anthropic's newest Fable and Mythos models, released the day before this story, came with lowered costs for cached tokens, and OpenAI cut its API prices at the end of July. A tier that lets a startup buy inference for a tenth of the list price, in exchange for something that costs it nothing on any single call, is a new form of price competition among the top model providers.

Source

Reporting: TechCrunch on the Meta contributor pricing tier. Primary source: Meta Model API pricing and rate limits.