Meta Offers Up to 98.7% API Discounts in Exchange for Muse Spark AI Training Data
Meta has introduced a massive data-sharing discount for its Muse Spark 1.2 and 1.3 AI models, cutting API costs down to between 1.3% and 8% of standard rates for users who consent to sharing their prompts and generated outputs for future model training. While most AI providers allow users to opt out of data collection, Meta is explicitly attaching a monetary value to prompt data to alleviate its training data shortage. This aggressive pricing strategy could establish a new industry model for incentivizing developers to supply valuable real-world interaction data. Under the contributor pricing tier, input costs fall from $1.25 to $0.10 per million tokens, output costs drop from $4.25 to $0.20 per million tokens, and prompt cache hit input costs decrease from $0.15 to $0.002 per million tokens. The initiative follows earlier difficulties in sourcing training data, including an internal employee tracking program that was shut down in June due to protests.
## BACKGROUND
As frontier artificial intelligence models grow in size, developers increasingly face a shortage of high-quality human data needed for training. Additionally, prompt caching is an optimization technique used in large language models that reuses pre-computed attention states from long context prompts, significantly reducing processing latency and computational expense on repeated requests.