AI's price collapse: the cost of matching top models is falling up to 900-fold a year
New research from Epoch AI finds the price of running large language models has plunged since 2022, with the sharpest falls arriving in the past year. The pace, however, varies wildly from one task to the next.
The cost of using artificial intelligence is falling at a speed with few parallels in modern technology. According to new analysis from the research institute Epoch AI, the price of running a large language model capable of a given task has dropped by anywhere between 9 and 900 times a year, depending on the capability in question.
The headline figure is striking on its own. To take one example, the price of reaching the level GPT-4 achieved on a set of PhD-standard science questions has fallen roughly 40-fold every year since the model's release. A capability that cost a small fortune to access in early 2023 now costs a tiny fraction of that, and the trend shows little sign of stopping.
Epoch AI, working with data from the independent firm Artificial Analysis, examined the results of leading models on six widely used benchmarks over the past three years, covering general knowledge, advanced mathematics, coding and graduate-level science. For each performance milestone, the researchers tracked the cheapest model able to match it, then measured how quickly that price fell.
The result was a median decline of around 50 times a year across all the tasks studied. But the spread was enormous. At the slower end, prices fell by a factor of nine annually. At the faster end, they collapsed by a factor of 900. That unevenness matters, because it means the economics of AI depend heavily on exactly what a business is trying to do with it.
Perhaps the most consequential finding concerns timing. The very fastest price drops, Epoch AI says, all began after January 2024. When the researchers stripped out data from before that date, the median rate of decline more than tripled, rising from 50 times a year to 200 times a year, even for capabilities that models had already reached before 2024.
That points to a broad acceleration over the past eighteen months rather than a handful of one-off bargains. It also carries a warning. Because the steepest falls are so recent, the researchers caution that it is far from certain they will continue at the same pace.
The report is careful not to overstate what it can explain. Some drivers are well understood. Models have become smaller and more efficient, and the hardware used to run them has grown steadily cheaper and more powerful. Fierce competition between providers has almost certainly played a part, as has the arrival of capable open-weight models that anyone can host.
Other factors are harder to pin down from public information. The researchers considered whether providers are simply accepting thinner profit margins to win market share, but said they found no clear evidence to confirm it. The prices are tumbling, but the full recipe behind the fall remains partly hidden.
Epoch AI is a research organisation that studies the long-run trajectory of artificial intelligence, describing its mission as investigating the technology for the benefit of society. It has built a reputation for tracking the hard numbers behind the AI boom, maintaining public databases on models, computing power, hardware and the data centres that train the largest systems.
The group is perhaps best known for FrontierMath, a notoriously difficult mathematics benchmark designed to test the reasoning limits of the most advanced models. Its regular data insights, of which this pricing analysis is one, have become a reference point for researchers, investors and policymakers trying to make sense of how quickly the field is moving.
Measuring intelligence by benchmark is an imperfect science, and the authors are candid about the limits of their work. Benchmarks do not capture everything that makes a model useful, and scores can be inflated when models are trained, deliberately or otherwise, on material close to the test itself. Results also wobble because of randomness in how models respond and small differences in how questions are put to them.
The analysis set aside so-called reasoning models, which work through a problem in many more steps and generate far more text. Comparing them on a simple price-per-token basis would flatter older models and mislead. The researchers note, however, that when they looked at the total cost of running full evaluations rather than the price of individual tokens, the downward trend held.
For businesses and consumers, the direction of travel is unambiguous even if the exact slope is not. The intelligence that sat behind an expensive frontier model two years ago is now available for a small fraction of the price, and often from more than one supplier. Whether the past year's extraordinary pace can be sustained is the open question. If it can, the assumptions built into a great deal of today's AI planning, about what is affordable and what is not, may need rewriting sooner than expected.