Standard measures miss 87% of the real price drop in AI services.

Posted prices for running artificial intelligence models have dropped steadily since 2024. However, normal measurement tools miss how fast this drop actually is. Statistical agencies usually track standard software by matching identical products over time. This approach misses models getting smarter. Researchers solved this by building a quality-adjusted price index. They linked public prices directly to AI test scores, creating a scale that rates models on how well they perform. They found that standard matched methods show prices falling at 0.10 log points a year, while their quality-adjusted measure dropped at 0.73. This means existing tools miss 87% of the real price drop. Yet, the price buyers pay per finished task actually stopped falling. Newer reasoning models use up words, or tokens, faster than token costs fall. Finally, cleaning up contaminated benchmark tests barely changed model ranks, but shifted the calculated price trend substantially.

Why it matters

Policymakers and economists rely on price indices to measure inflation, productivity, and market competition. Because standard methods miss 87% of the real decline in AI inference costs, they overstate inflation and undercount productivity improvements. For buyers, the per-task cost has stopped falling entirely because reasoning models burn through tokens faster than token prices drop, shifting the real cost of using AI from compute to reasoning overhead.

Questions answered