The inference-cost wedge and free-tier quality in generative AI freemium

Generative AI platforms price free tiers at zero yet face per-query inference costs that rise with model capability. We study a freemium monopolist offering free and paid tiers. Decomposing the free-quality first-order condition reveals an inference-cost wedge beyond standard screening: serving costs for existing free users, and a net-of-serving-cost entry margin. Under quadratic costs, optimal free-tier quality is strictly decreasing in inference costs whenever a closed-form sufficient condition holds. Because the free-tier price is fixed at zero, the free tier responds to cost changes through quality rather than price, so falling inference costs raise optimal free-tier quality.