How fast is AI progressing?

The speed of AI can be measured in several ways: the price of the same service, the price of the hardware that delivers it, what the models can do, the amount of computing put into them. To compare prices that fall at very different rates, this page expresses each as the time it takes to be divided by 10, with 2 older benchmarks: light, whose price the economist William Nordhaus reconstructed over 2 centuries, and the computing of graphics cards. These rates are our calculations, from the published series.

Epoch AI collected, for 6 tests and several score thresholds, the lowest price at which a model reaches that threshold, month after month. The fall ranges from 9 to 900 times a year depending on the threshold. To reach the GPT-4 score on a test of doctoral-level science questions, the price fell by a factor of 40 a year. Epoch AI notes that the fastest falls are from the last year and that it is less certain they will last.

The Stanford AI Index gives an example: a model as good as GPT-3.5 on the MMLU general knowledge test cost 20 dollars per million tokens in November 2022, and 0.07 dollar in October 2024, 280 times less.

These prices are those the providers list. They follow the cost of the service, but also competition and commercial choices; they do not say how much energy is used.

Hardware progresses much more slowly. Epoch AI measures that the number of operations per second bought with one dollar of graphics card doubles every 2.5 years, and that the number of operations per watt doubles every 2.4 years since 2008: it takes about 8 years to divide the electricity of the same computation by 10. For the most recent AI chips, computing per dollar doubles every 1.7 year since 2023.

Hardware therefore explains only part of the fall in the price of responses; the rest comes from the models, the serving software and the prices set by the providers. Our analysis of AI decarbonisation measures these levers, and our explanation on graphics cards says why memory limits the throughput of a card.

The METR organisation measures capabilities by the duration of tasks that a model completes one time in 2, counted as the time a professional would spend on them. On software and machine learning research tasks, this duration doubles about every 7 months since 2019. In its January 2026 update, METR finds a doubling in 131 days, about 4 months, for models released since 2023, with a range of 107 to 161 days, and in 89 days since 2024. METR writes that the uncertainty of its model is large, and says it is fairly confident that the trend lies between 1 and 4 doublings a year. The tasks are mostly software tasks; the measurement says nothing about other occupations.

The amount of computing used to train the largest language models is multiplied by 5 every year since 2020, according to Epoch AI, a doubling about every 5 months. It grows faster than the energy efficiency of cards, which is multiplied by about 1.3 a year: the electricity used to train the largest models increases from one generation to the next, according to our calculation from these 2 rates.

At Google, the energy of a median Gemini query was divided by 33 in one year, while the text processed was multiplied by 50; our analysis of AI consumption over time follows these 2 series and Google's total electricity.