The Emerging Market for Intelligence: Pricing, Supply, and Demand for LLMs
Fradkin Demirer, Peng Tadelis
Cost per token at the offered price is stable But intelligence per token is up (→ token per intelligence is down) So cost per unit of intelligence is down Weighted token price is actually up in 2025 (people shifted toward higher quality / more expensive models)
This is classic Simpson’s Paradox – across models price and quantity might actually be positively correlated, but within model they are negatively correlated
Also interesting that high throughput only matters without the model fixed effects, but once you have those it basically goes to zero. This implies that higher throughput models are used more (across models) but within model throughput is not important to usage patterns (or even slightly negative)
Latency is like price – it matters more when you control for model, which means that in the unconditional data latency is positively correlated with quantity (high latency models are used more)
Context length is like price – it’s positively correlated with quantity, but that is greatly reduced within model. It’s still a positive even within model though, but it’s a noisy relationship
Interesting price trends, shows that most of the decline in price is from new cheaper model rather than old models getting cheaper. Also shows that a given closed source model doesn’t really get cheaper over time, only open source models get cheaper over time
Can also see this visually:
Could do something like @melitzDynamicOlleyPakesProductivity2015 to break down price changes between new models and existing models getting cheaper.
Also interesting that token weighted prices actually increased during 2025. Also points out that closed source models rarely change their pricing:
Pricing across categories much higher variance than intelligence by category, suggesting intelligence can’t explain. Rank correlation also appears to be moderate
