New research from the institute Epoch AI indicates that the financial barrier to using advanced artificial‑intelligence models is eroding faster than any comparable technological trend in recent history. By tracking the cheapest system capable of matching specific performance milestones across six widely used benchmarks, the study finds that prices are falling anywhere between nine and nine hundred times each year, depending on the task.
Scale of the decline
The median rate of price reduction is roughly fifty‑fold annually. The distribution, however, is extremely uneven: at the low end, costs drop by a factor of nine per year, while at the high end they shrink by as much as nine hundred times annually. This variance matters because the economics of AI adoption hinge on the exact capability a business requires.
One striking illustration comes from the cost of achieving the level of performance that GPT‑4 demonstrated on a set of PhD‑standard science questions. Since the model’s launch, the expense of reaching that benchmark has declined by about forty times each year, turning what was a “small fortune” in early 2023 into a modest outlay today.
Accelerating after 2024
The most rapid price drops all began after January 2024. When the researchers excluded data from before that month, the median annual decline more than tripled, rising from fifty‑fold to two hundred‑fold. This suggests a broad‑based acceleration over the past eighteen months rather than isolated discount events.
Despite the recent steepness, the authors caution that the pace may not be sustainable, noting that the steepest declines are a very recent phenomenon.
Drivers behind the falling costs
Several well‑understood factors contribute to the trend. Model architectures have become more compact and efficient, while the hardware required to run them has steadily become cheaper and more powerful. Intense competition among providers and the emergence of open‑weight models that anyone can host have also played a role.
The study considered the possibility that providers are accepting slimmer profit margins to win market share, but found no clear evidence to support that hypothesis.
About Epoch AI
Epoch AI describes its mission as studying the long‑term trajectory of artificial intelligence for societal benefit. The organisation maintains public databases on model performance, computing power, hardware, and data‑centre usage. It is perhaps best known for the FrontierMath benchmark, a notoriously difficult test of mathematical reasoning that has become a reference point for researchers, investors and policymakers.
Caveats and limitations
Measuring intelligence through benchmarks is an imperfect science. Scores can be inflated if models have been trained on material similar to the test, and random variation in model responses can affect results. The analysis deliberately excluded “reasoning models” that generate long, multi‑step outputs, because a simple price‑per‑token metric would unfairly advantage older systems. Nevertheless, when the total cost of running full evaluations was examined, the downward trend persisted.
For businesses and consumers, the direction is clear: capabilities that required expensive, frontier‑level models two years ago are now available for a fraction of the price, often from multiple suppliers. Whether the extraordinary pace of the past year can be maintained remains an open question, and could force a rapid reassessment of current AI budgeting assumptions.
First reported by Epoch AI.
