According to the Epoch AI report, the cost of achieving a specific level of AI performance has been dropping by an average of 47% per quarter since 2023. This equates to a 13-fold annual reduction, outpacing historical trends in DNA sequencing, computing power, and battery technology.
The research highlights that costs for cutting-edge AI capabilities fall fastest when they first reach industry-leading status, averaging a 66% quarterly decline. After two years, this rate slows to approximately 32% per quarter, suggesting that the competitive advantage of frontier models is shrinking rapidly.
Cost reductions vary by domain, with mathematical reasoning capabilities becoming cheaper by 50% to 52% per quarter. Tasks involving games and puzzles see a slightly lower quarterly decline of 39% to 43%. This trend is driven by hardware efficiency, model architecture improvements, and intense market competition.
Despite the overall price collapse, the report notes that the absolute cost of running top-tier models may remain high. Developers continue to invest in more computational resources to push for higher accuracy and reasoning power, even as the cost for a fixed performance level plummets.
