What the paper measured
Buying a fixed level of AI performance has become about 47 per cent cheaper every quarter since 2023 — roughly 13 times cheaper a year — according to an analysis published on 22 September by Luke Emberson and David Roodman at Epoch AI.
The method matters more than the headline. Rather than track what the average model costs, the authors track the cost frontier: for each level of performance, the cheapest model that reaches it at any point in time. That turns a messy field of releases into a single falling line, and it is the line a buyer actually faces, since nobody is obliged to keep paying for last year’s most expensive option.
The estimate covers five benchmarks spanning mathematics, the hard sciences and games of skill.
Not every task falls at the same rate
The decline is uneven. Maths problems get cheaper fastest, at 50 to 52 per cent a quarter. Game-based puzzles are the slowest in the set, at 39 to 43 per cent.
There is also a pattern in time. Performance that has just arrived at the frontier gets cheap very quickly — the authors put it at 66 per cent a quarter, about 75 times a year — and then the rate halves. Two years after a capability first appears, its price is falling at 32 per cent a quarter, or 4.7 times a year. The expensive thing is being early.
The comparison that gives the number its weight
Epoch’s point is comparative. By its figures the AI cost curve runs about four times faster than DNA sequencing, six times faster than computing itself, 18 times faster than lithium batteries, and 54 times faster than electricity over the century to 1973.
The caveats are in the paper
The authors list the obvious objections themselves. Models may be trained with the benchmarks in mind, which flatters the numbers. Benchmarks are imperfect stand-ins for what people actually want done. Real users do not switch to the cost-optimal model the moment one appears, so the frontier price is not the price most people pay. Three years is a short series to fit a trend to. And reasonable alternative choices in the method give quarterly declines anywhere between 42.9 and 58.0 per cent.
That range is worth holding on to. The claim that survives it is directional rather than precise: the cost of a given capability is collapsing, faster than the historical analogies suggest it should, and the part of the bill that is not falling is the part spent at the frontier.
What to watch
The number to check next is whether the rate holds as the newest reasoning models, which spend far more compute per answer, work their way into the frontier. A cost curve built on cheap inference behaves differently when the cheapest model that reaches a score is one that thinks for a long time to get there.
