The plunging price of thought

Source: Epoch AI (Luke Emberson, David Roodman) · 2026-09-22

AI has gotten cheaper more quickly than any other transformative technology in history. The cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year. That price drop is 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.

Coarser evidence suggests that the price of thought has been falling at least this fast since the dawn of commercial LLM inference in November 2021, when OpenAI fully released GPT-3.

The speed of price drop varies by domain: slower on game-based puzzles, at 39–43% per quarter, faster on math problems, at 50–52% per quarter.

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). We see this pattern on three of our five main benchmarks of AI capability. Averaging across all 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).

It is now widely understood that the AI boom is a macroeconomic force powerful enough to raise prices for the inputs it demands: chips, power, even the labor of electricians. Less well recognized is a paradoxical flip side: the price of the output from all those data centers is falling extraordinarily rapidly:

The next chart shows some examples. On January 31, 2025, OpenAI released a new iteration in its series of “reasoning” models, called o3. We estimate that for an average cost of 30 cents per question, it could achieve a 75% score on GPQA Diamond, a multiple-choice exam covering PhD-level physics, chemistry, and biology.1 Just under 18 months later, OpenAI released GPT-5.6 Luna. It scored just as well — for four hundredths of a penny per question ($0.0004). That is a 725-fold drop in the price of thought in under 18 months. It is like the sticker price on a new car falling from $50,000 to $69. No other general-purpose technology in history appears to have gotten so cheap so fast.

Across that time, we find that the price for a given level of performance has fallen about 47% per quarter, or 13× per year. We see slower drops on game-based puzzles, at about 39–43% per quarter (7–10× per year), and faster progress on math problems, at 50–52% per quarter (16–19× per year).

The cost decline for a given level of performance does tend to slow over time, though this pattern is not universal. One possible explanation: 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. Averaging across the five primary benchmarks, cost falls 66% per quarter (75× per year) at first. Two years later, it falls half as fast, at a “mere” 32% per quarter (4.7× per year).

Our analysis comes with major caveats. AI companies may be expressly training their models for some benchmarks (“benchmaxxing”), so that improvement on the benchmarks outstrips improvement for real-world tasks. Even if they are not, doing well on a benchmark is not synonymous with useful work. Because we focus on the frontier — the absolute cheapest model capable of any given level of performance — we implicitly posit an AI user who relentlessly searches for the most cost-effective model for each task, when real users do not switch models so often, and therefore do not reap quite the same savings. Our data are incomplete and noisy: the timeframe is barely three years, and we do not include all combinations of AI model and benchmark. Prices drop differently for different models, benchmarks, time periods, and performance ranges, and there are many reasonable ways to average over this variegated experience. Overall, while we believe that our bottom-line numbers are reasonably representative of reality, they should not be read as exact.

All of these analyses measured prices per token for models capable of achieving a given performance as distinct from the actual cost to achieve that performance. With the advent of reasoning models — OpenAI released o1 in December 2024 — it has become more problematic to ignore this distinction. Reasoning models can productively consume far more tokens, but as a result extract good performance from an underlying LLM that is smaller and cheaper to run per token.

Among the five primary benchmarks, the decline rates are fairly close, ranging from 43% per quarter for Chess Puzzles to 53% per quarter for FrontierMath, tiers 1–3. Their simple average is 47% per quarter.

In three of the five primary benchmarks, and all of the secondary ones, the rate of cost decline falls after SOTA debut. In the exceptions — Chess Puzzles and Mystery Game Puzzles — the acceleration amounts to only a few percentage points. On average for the primary benchmarks, the decline rate drops from 66% per quarter on average at SOTA to 32% two years later.

While the fundamental finding of rapid price decline is indisputable, our particular estimates are subject to major caveats.

First, there could be “benchmaxxing”: the deliberate training of AI models to perform well on known benchmarks. One of our primary benchmarks, Mystery Game Puzzles, is designed to mitigate benchmaxxing by keeping the game in question secret. In Table 1, the quarterly decline rate for this benchmark lies somewhat below the average of the five, at 44.0% (model-free) or 38.7% (Box-Tidwell model for frontier) versus 47.0%. These results are consistent with the benchmaxxing critique being valid but not fatal.

Second, benchmarks are an imperfect proxy for useful work. Thus our cost figures are imperfect proxies for the market value of LLM inference as a service. Just below, we will compare the trend in the price of thought to trends for other transformative technologies, such as electricity.