Meta launched a state-of-the-art open model, Muse Spark, alongside a new two-tier API pricing system that makes an explicit data-for-access barter. The Standard Tier (muse-spark-1.3) charges $1.25/m input tokens and $4.25/m output tokens, with a guarantee that prompts and completions are never used to train future foundation models. The Contributor Tier (muse-spark-1.3-contributor) charges $0.10/m input and $0.20/m output — a 92–95% discount — in exchange for Meta retaining the right to train on the data. No other foundation model provider currently offers this explicit trade on its API.

The article frames this as the advertising business model arriving in AI. At a 30:1 agentic input-to-output ratio, the blended standard cost is ~$1.35/m tokens; the contributor cost is ~$0.103/m. The $1.24/m spread represents the implicit value Meta places on customer data. For an enterprise processing 1 billion tokens per day, opting for the private tier translates to a $454,000 annual “privacy surcharge.” The pricing unbundles the flat-rate consumer subscription model (e.g., $20/month) where training rights were bundled with access. Meta has now formalized the data subsidy per token.

Why surrender 92% of inference revenue? The argument is not altruistic. The public web is exhausted as a training source; frontier gains now come from post-training, RL with AI feedback, and usage patterns. The market for labeled training data is estimated at $10B annually. By offering deeply subsidized inference in exchange for organic reasoning traces, Meta bypasses labeling vendors and vertically integrates the data supply chain — just as search and social platforms did with behavioral data for advertising. Compute is no longer a utility; it is currency traded for future training tokens. This explicitly provides a business model for American open-source AI: subsidized access in exchange for data, analogous to ad-supported services.