Inference costs have dropped roughly 40% year-over-year, and the trend is accelerating. For builders, this shifts what's economically viable to build.
Two years ago, running a capable model in production required careful budgeting and aggressive caching. Today, the same workload costs a fraction of what it did — and the providers keep competing on price. That's good news for anyone shipping AI-native products, but it also changes the competitive landscape in ways founders should think about carefully.
The data
Across the three largest API providers, the median price per million tokens for frontier-class models has fallen steadily. Hardware efficiency improvements, better serving stacks, and fiercer competition have all contributed. When inference gets cheaper, the moat moves from "who can afford to run this" to "who has the best product on top of it."
What founders should do
First, re-run your unit economics. Products that were margin-negative at last year's prices may be viable now. Second, assume your competitors are doing the same calculation — features that felt like luxuries will become table stakes quickly.
Third, watch the open-weight models. When the cost of a capable baseline approaches zero, differentiation comes from data, distribution, and workflow design rather than raw model access.
The bottom line
Cheaper inference is a tailwind for every company building on top of these models. The startups that win will be the ones that treat it as an opportunity to rebuild their product economics — not just to cut their cloud bill.