The Plunging Price of Thought: AI Performance Costs Fall 47% per Quarter

Summary

Epoch AI estimates that the cost of achieving a given level of AI performance has fallen by about 47% per quarter over the past three years, equivalent to a 13-fold annual drop. The estimate is based on five primary benchmarks covering mathematics, hard sciences and games of skill. Costs fall more slowly for game-based puzzles, at roughly 39–43% per quarter, and faster for mathematics, at 50–52%. The report also finds that prices often decline fastest soon after a performance level first becomes state of the art: across the five benchmarks, the average decline is about 66% per quarter at debut, slowing to 32% per quarter two years later. As an example, Epoch estimates that OpenAI’s o3 could reach a 75% GPQA Diamond score for about $0.30 per question, while GPT-5.6 Luna achieved the same score for $0.0004, a 725-fold reduction in under 18 months. The analysis constructs cost-performance Pareto frontiers from model evaluations, including different reasoning levels and spending limits, and uses frontier models to estimate trends. The authors caution that benchmark optimization, imperfect data, limited real-world model switching and the short three-year timeframe make the figures approximate. Falling inference prices may not eliminate total AI spending because demanding tasks can require very large numbers of runs, while capability continues to rise.