Freemium AI pricing: New paper proves free tiers optimize revenue and quality

Researchers Maxim Bichuch (Johns Hopkins) and Aviv Yaish (University of Waterloo) have released a 39-page paper, "Freemium Is All You Need," that tackles a central question for GenAI companies: Should you give away free access? The paper, posted on arXiv (2608.00823), models a service provider whose users differ in request value and privacy preference. Paid requests stay private, while free requests generate data that improve future model quality. Without that data, quality reverts toward a baseline—a clever way to capture model drift.

The authors prove that for any stationary service strategy, quality converges to a unique steady state, meaning companies can predict long-term model health. More importantly, they derive optimal pricing and quality policies expressed using two endogenous user value thresholds. For uniform value distributions, they provide a closed-form optimal strategy and characterize sufficient conditions for offering free services based on inference cost. Essentially, if the cost of serving a free request is below a calculated threshold, the data it generates is worth more than the compute expense—making "free" a rational profit-maximizing choice. The paper bridges economic theory and ML operations, offering a rigorous mathematical foundation for the freemium models already used by OpenAI, Anthropic, and Google, but with optimized thresholds instead of intuition.