Epoch AI says cost of a fixed benchmark score falls 13x a year
Two teams measured how fast the price of reaching a given score drops, and they differ on how much of the fall comes from algorithms.
Epoch AI said the cost of reaching a fixed score on AI benchmarks is falling about 47% a quarter, or roughly 13 times a year. It measured five benchmarks covering maths, science and logic puzzles.
Its example is GPQA Diamond. Reaching o3's 75% accuracy cost about 30 cents a question in early 2025, and about 0.04 cents 18 months later, Epoch AI said. That is a fall of about 1/725.
A separate MIT group reached a smaller figure for the part that comes from algorithms. Hans Gundlach and colleagues used Artificial Analysis data from April 2024 to November 2025 and measured costs falling 5 to 10 times a year, or about 3 times once they stripped out hardware gains and price competition.
Neither figure means today's best models are cheaper to run. Reasoning models spend far more compute per task, so a higher benchmark score may come from buying more processing rather than from efficiency. The Decoder is so far the only outlet to report the two sets of numbers together.
Epoch AI said that "no other transformative technology declined that fast". It describes its own figures as "reasonable but rough measurements based on the best available data", given the narrow set of benchmarks behind them.