How fast is AI capability becoming cheaper to produce?

Epoch AI (Luke Emberson, David Roodman) estimate that "the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year" — a decline they describe as faster than any other transformative technology in history, "four times faster than DNA sequencing, six times faster than compute, 18 times faster than lithium batteries, and (in the century up to 1973) 54 times faster than electricity." The rate varies by domain: "slower on game-based puzzles, at 39–43% per quarter," and "faster on math problems, at 50–52% per quarter." Their starkest illustration: OpenAI's o3, released January 31, 2025, scored 75% on GPQA Diamond (a PhD-level physics, chemistry, and biology exam) for an estimated 30 cents per question; "just under 18 months later," GPT-5.6 Luna matched that score for $0.0004 per question — "a 725-fold drop in the price of thought," which the authors compare to "the sticker price on a new car falling from $50,000 to $69."

Epoch AI's method tracks "the absolute cheapest model capable of any given level of performance" across five primary capability benchmarks (including Chess Puzzles, Mystery Game Puzzles, and FrontierMath tiers 1–3), rather than tracking any single model's price over time. Coarser evidence, they write, suggests the same pace of decline reaches back "to the dawn of commercial LLM inference in November 2021, when OpenAI fully released GPT-3." They set the finding against a macroeconomic backdrop in which "the AI boom is a macroeconomic force powerful enough to raise prices for the inputs it demands: chips, power, even the labor of electricians" — even as the price of the output those inputs produce falls sharply.

This prices AI capability in dollars at a different layer than How much progress have AI agents actually made on NanoGPT? and When do AI agents outperform human research experts?, which price an agent's R&D contribution against a human baseline rather than tracking how inference serving cost for a fixed benchmark score falls over calendar time; together the three give complementary dollar-denominated views of AI economics — serving cost (here), optimization value, and comparative task cost — without one superseding another. Do AI performance prices fall faster when capabilities debut? (sibling note) supplies the shape of this same decline curve that the 47%-per-quarter average flattens into a single number.

Epoch AI names its own caveats: "benchmaxxing" (deliberate training toward known benchmarks) could inflate benchmark gains relative to real-world task gains; the "frontier" measure assumes "an AI user who relentlessly searches for the most cost-effective model for each task," which real users don't do, so realized savings are likely smaller; and the dataset spans "barely three years" with incomplete model and benchmark coverage, so the authors say their bottom-line numbers "should not be read as exact." The finding establishes a rapid and historically unusual price decline for benchmark-measured AI performance; it does not establish that real-world AI costs for typical users fall at the same rate, nor that benchmark score gains track real-world usefulness at a fixed ratio.