AI performance costs are falling faster than those of any previous technology

Research by Epoch AI finds that the cost of reaching a fixed performance level on AI benchmarks has fallen about 47 percent per quarter since 2023, roughly 13x per year, faster than any previous transformative technology. MIT researchers put the annual decline at 5x to 10x, with about 3x per year attributable to algorithmic efficiency after excluding cheaper hardware and competitive pricing. Epoch cites OpenAI's o3: in early 2025 it scored 75 percent on the GPQA Diamond science test at an estimated 30 cents per question; eighteen months later, a GPT-5.6 family model reached the same score for four hundredths of a cent, 1/725 of the original price. However, today's best reasoning models can still cost more per query because they use far more compute per task.