A recent publication from Epoch AI, authored by Luke Emberson and David Roodman and titled “The Plunging Price of Thought”, estimated that the cost of achieving a fixed level of artificial intelligence (AI) performance has fallen by roughly 47% each quarter, about a 13-fold decline each year. The paper measures the price of thought as a commodity. It asks: For a given level of measured performance, what is the cheapest available model capable of achieving it? This is closer to a quality-adjusted technological price frontier.
The authors estimated that AI prices fell four times faster than DNA sequencing, 18 times faster than lithium-ion batteries, and 54 times faster than electricity. These results come from five benchmarks that cover mathematics, hard science, and game skills. The pace varies across benchmarks, with costs falling roughly 39-43% per quarter for game-based puzzles and 50-52% faster for mathematics problems. Overall, the findings suggest no general-purpose technology has become so cheap so quickly.
The research finds that frontier performance depreciates fastest when it first debuts as the state-of-the-art across five benchmarks; capability costs fall about 66% per quarter initially, slowing to 32% after two years. This pattern is pronounced in three benchmarks. But benchmark performance is not synonymous with intelligence. AI models may be deliberately optimised for widely used benchmarks, a practice the paper calls “benchmark-maxxing”. The better results we are seeing across the tasks they are trained on could reflect improved training data, post-training test-time reasoning, or greater general capability, in varying proportions. That said, benchmark-maxxing alone cannot explain the result, nor is benchmark performance a clean measure of an underlying quality called intelligence.
One plausible explanation, the authors say, is that the new capability achieved by large language models (LLMs) commands a scarcity premium before competition and engineering improvements rapidly reduce costs. This suggests rapid technological progress, but it doesn’t necessarily mean the cost of producing intelligence itself is falling at the same rate.
{{/usCountry}}One plausible explanation, the authors say, is that the new capability achieved by large language models (LLMs) commands a scarcity premium before competition and engineering improvements rapidly reduce costs. This suggests rapid technological progress, but it doesn’t necessarily mean the cost of producing intelligence itself is falling at the same rate.
{{/usCountry}}Frontier economics is moving in the opposite direction. Separate research from EPoch suggests that frontier language-model training costs have risen about 3.5 times annually since 2020, even as reproducing a fixed level of capability has become cheaper. Yesterday’s capabilities may cost less to deliver while labs spend more to advance, but cheaper capabilities don’t guarantee lower user prices or greater access.
Users ook at benchmark capabilities, but in the end buy the overall package and what it offers, i.e., the product. As the cost of reproducing LLM capability falls, the cost of developing the frontier can rise, and the price consumers pay can follow a different path altogether. This matters most for countries outside the highly concentrated AI frontier. For most countries, the immediate question may be how to absorb technology whose architecture, commercial trends, and technological directions are largely determined elsewhere. This changes the question, especially for economies in the Global South: If AI cognition becomes dramatically cheaper, which domestic factors of production determine whether countries can actually turn it into productivity? One of them is human capability. The technology does not enter an economy in isolation; its effect depends on complementary institutions, infrastructure, and human skill.
AI models’ capability has an unusual property as a factor of production. Once created, it can be copied, distilled, and deployed at enormous scale. However, can we say the same for human capability? Developing a capable scientist, engineer, teacher, or researcher still requires decades of education, practice, experience, and rigorous training. India illustrates this challenge. PARAKH’s 2024 assessment, covering over 2.1 million students in grades 3, 6, and 9, points to weaknesses in the foundational competencies at later grades, especially mathematics. When these competencies are at their weakest, importing powerful insights from LLMs does not automatically teach people to question, verify, or build on their outputs. AI may replace some forms of human cognition while increasing the value of others: identifying worthwhile questions, forming hypotheses, exercising judgment under uncertainty, conducting experiments, and creating new knowledge to understand real-world use cases. For economies behind the AI frontier, falling costs shift the challenge from model access to building the human and institutional capacity to boost productivity.
We do not yet know whether producing a capable human mind will itself become more expensive, but in a world of abundant, cheap AI, producing thought through AI models is cheap. The economic cost of failing to produce capable human minds could be much higher.
Amit Kapoor is chair, and Sheen Zutshi is research manager, Institute for Competitiveness. The views expressed are personal