Buying a fixed level of AI ability gets about 13× cheaper each year, Epoch AI found
The price of answering at a fixed standard halves about every quarter: one science level got 725× cheaper in 18 months. Top models did not.
The usual price measure, cost per token, flatters newer models, because they “think” for thousands of extra tokens before answering and the user pays for those too. Epoch AI’s September 2026 study instead measured the cost of actually reaching a given score, reading the transcripts of five test sets covering maths, science and puzzles.
The price of a fixed level of ability fell by about 47% per quarter, roughly 13× a year. Maths fell fastest, game puzzles slowest. In Epoch’s own example, o3 answered PhD-level science questions at about $0.30 each, and GPT-5.6 Luna later answered them just as well for $0.0004, a 725-fold drop in under 18 months.
The fall is fastest right after a level is first reached, about 75× in the following year, and slows to about 4.7× a year once the level is two years old.
What it means
If a task needs a level of ability that AI reached a year or two ago, running it costs a small fraction of what it did, which is why features that were once too expensive to ship now appear in ordinary apps. This says nothing about the price of the best model available: that is a fixed standard getting cheaper, not the frontier. Waiting a year is the cheapest way to buy any given level of ability.
Caveats
- Test scores can be inflated if similar questions were in a model’s training data. The one test whose questions are kept secret showed a slower fall.
- The figures assume users always switch to the cheapest model that is good enough, which few do.
- There are only about three years of data, and averaging it a different way gives 43–58% per quarter.