Meta offers ~95% price discount on new Muse Spark model API for customers who let it train on their prompts and outputs

Meta has reportedly introduced a two-tier pricing structure for its new Muse Spark model API. Standard pricing is $1.25 per million input tokens and $4.25 per million output tokens, while a "contributor" tier that requires sharing prompts and outputs for future model training drops the price to $0.10 per million input tokens and $0.20 per million output tokens, a discount of roughly 92-95%. The report rests on TechCrunch's account rather than a linked primary source such as Meta's own pricing page, and Meta reportedly declined to comment.

This is a concrete, actionable pricing decision: developers and enterprises evaluating Muse Spark now face a real trade-off between cost and data privacy, with numbers specific enough to act on. It also signals Meta sees a shortage of real-world usage data, particularly for agentic and professional workflows beyond coding, as a genuine bottleneck worth paying steeply for. The catch is that the enterprises most likely to hold the most valuable, differentiated data are also historically the most resistant to trading it away even at a steep discount, so the scheme's actual pull on the customers Meta most wants remains unproven.

The framing that this "reveals" a broader crisis in enterprise data access leans heavily on one expert's opinion and Meta's own marketing language, not on disclosed uptake figures. Anthropic and OpenAI are name-checked for context on price competition, but there is no evidence either has done anything similar, so treat the "broader frontier-lab trend" framing as speculative rather than established.

1. 01Whether Meta publishes this pricing on its official developer or pricing documentation (currently unconfirmed, no primary source linked) and how it precisely defines contributor-tier data use in its terms
2. 02Actual enterprise uptake of the discounted tier over the next one to two quarters: does anyone with valuable agentic or professional-workflow data opt in, or does uptake stay confined to hobbyists and low-stakes users
3. 03Whether Anthropic, OpenAI or other labs introduce similar data-for-discount tiers, which would confirm this is a real industry pattern rather than a one-off Meta move