AI Gets 13x Cheaper Each Year — Faster Than Any Other Technology

Epoch AI has calculated that the cost of getting an identical result from language models falls by 47 percent per quarter on average. That works out to roughly 13 times a year. This is faster than any other technology in history has improved. Here is a concrete example. Solving the same task on the GPQA Diamond benchmark cost 30 cents with the o3 model in early 2025. Eighteen months later, a model from the GPT-5.6 family matched that result for just four hundredths of a cent. By Epoch's math, that is 1/725th of the original price. Researchers at MIT checked these numbers separately, filtering out the effect of cheaper hardware and competition between companies. They arrived at a different figure: real gains in algorithmic efficiency run at about 3 times a year.

This drop has a limit. Reasoning models spend more compute on each individual task. MIT also points out that not all benchmark progress reflects real efficiency gains. Some of it just comes from throwing more compute at the problem. The hardest runs, where an agent needs to reason for a long time, are not getting cheaper. Sometimes they get more expensive. I work with Claude every day, and from my own experience I see the same pattern. The savings are real on routine tasks. They are much less visible where the model has to think for a long time.