A study by Epoch AI finds that the cost of AI performance falls thirteenfold every year.

Epoch AI compared the cost of reaching set benchmark scores with AI models over time. The price per unit of performance drops thirteenfold per year, faster than DNA sequencing, computing power, or lithium batteries.

Research firm Epoch AI tracked the cheapest way to reach a given score on five performance benchmarks, covering math, science, and skill games such as chess. The price per benchmark point fell about 47 percent per quarter on average, a thirteenfold drop per year. For math problems the decline reached 16 to 19 times per year, for game-based tasks 7 to 10 times. One example: OpenAI's o3 model scored 75 percent on the GPQA Diamond test in January 2025 at a cost of 30 cents per question. About eighteen months later, GPT-5.6 Luna matched that score for 0.04 cents per question, a 725-fold drop.

That makes AI performance cheaper faster than any other technology Epoch studied: four times faster than DNA sequencing, six times faster than general computing power, and 18 times faster than lithium batteries. For companies, this means build-versus-buy decisions on AI have a short shelf life; a model considered expensive today can be a fraction of the price within a few quarters. Infrastructure budgets should plan for this pace rather than assume stable prices.

The authors note that benchmark scores are an imperfect proxy for real market prices. The price decline also slows over time: shortly after a new performance level is reached, cost falls about 75-fold per year, but only 4.7-fold two years later. It also remains open whether benchmark-only cost reflects the reality of production work, where review steps, rework, and human sign-off add cost and could flatten the curve.

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