Epoch AI tracked the cheapest way to hit fixed benchmark scores over time and found a 47 percent quarterly cost decline. One example: a 75 percent score on GPQA Diamond cost 30 cents per question in January 2025 and just 0.04 cents in mid-2026.
Epoch AI measured the cheapest way to reach a fixed score on five benchmarks covering math, hard sciences and strategy games, tracking how that cost changed over time. The average decline is 47 percent per quarter, equal to a factor of 13 per year. Math benchmarks fell fastest, at 50 to 52 percent per quarter, while game-based puzzles fell slower, at 39 to 43 percent. One example from the report: OpenAI's o3 model reached a 75 percent score on the physics and chemistry exam GPQA Diamond for 30 cents per question in January 2025. Just under 18 months later, GPT-5.6 Luna matched that score for 0.04 cents, a 725-fold drop.
Epoch AI says this makes AI the fastest-falling cost curve among all transformative technologies it has studied, four times faster than DNA sequencing, six times faster than general computing power, and 18 times faster than lithium batteries. For companies, that means today's cost assumptions for AI applications can become outdated within a few quarters. Anyone planning multi-year AI infrastructure budgets should build in the expected price decline rather than extrapolating current prices. The study also shows that cost drops fastest right after a new performance level is reached, then slows to roughly half that pace as competition and technical progress erode the initial premium.
How reliable these numbers are remains open. One commenter on the report notes that investor capital subsidizes part of the price decline, since AI providers absorb operating losses to win users, and that including those losses could flatten the curve considerably. Benchmark scores are also an imperfect proxy for actual market prices. Whether the trend holds once funding conditions for AI providers shift is a question only the coming years can answer.
What this means for decision-makers
- Check whether multi-year AI budgets already account for the expected drop in inference cost.
- Avoid long fixed-price AI contracts that ignore the expected pace of cost decline.
- Watch whether providers are subsidizing losses before treating current prices as stable.
This story was produced automatically from the source named above and checked by software before publication. The image is symbolic and shows neither the event nor a real person. How this paper is made