Do AI performance prices fall faster when capabilities debut?

Epoch AI find that within their five primary capability benchmarks, "the cost of a given level of performance often falls fastest right after that level is first achieved, that is, when it is state of the art (SOTA)." Averaged across the five, cost falls "66% per quarter (75× per year) for performance that has just debuted as SOTA"; "two years later, prices fall half as fast, at 32% per quarter (4.7× per year)." The pattern holds broadly: "In three of the five primary benchmarks, and all of the secondary ones, the rate of cost decline falls after SOTA debut." The two exceptions among the primary set, Chess Puzzles and Mystery Game Puzzles, drift the other way, but "the acceleration amounts to only a few percentage points." Per-benchmark decline rates otherwise cluster tightly, "ranging from 43% per quarter for Chess Puzzles to 53% per quarter for FrontierMath, tiers 1–3."

Epoch AI's proposed explanation is competitive rather than purely technical: "when a performance level is first achieved, AI companies can briefly charge a premium for it, before competition and technological improvement quickly drive down the price. In time, that dynamic slows." A newly-SOTA capability is priced, in effect, like a temporary monopoly; competing labs and newer, cheaper models erode that premium quickly at first, then more slowly as the performance tier matures and fewer competitors remain to undercut it further.

This refines the headline rate in How fast is AI capability becoming cheaper to produce? (sibling note): the 47%-per-quarter figure is a cross-benchmark average that flattens a two-speed curve — fast while a capability is new, slower as the tier ages — into one number. It is also a different cost lens than How much progress have AI agents actually made on NanoGPT?, which prices an agent's R&D optimization contribution as a one-time dollar threshold rather than following a price over calendar time, and than When do AI agents outperform human research experts?, which compares a fixed per-run agent cost to a human cost rather than tracking any price trend.

Epoch AI calls the SOTA-premium explanation "one possible explanation," not a settled mechanism, and notes the pattern "is not universal" — two of five primary benchmarks run the other way, even if only slightly. The same major caveats apply as to the headline rate: possible benchmaxxing, the frontier-search assumption that overstates realized user savings, and a data window of "barely three years." The finding supports a narrower claim than "AI gets cheaper over time": within these benchmarks, newness commands a premium that competition erodes on a roughly two-year-ish half-life, but whether that premium-decay dynamic generalizes to capabilities outside these five benchmarks is not established here.