Right after an AI model sets a new record, the price to match that performance crashes fast — then the price drops slow down a lot. Why?

This explores why the price of a given level of AI performance falls fastest right after a model first reaches it, and what is behind that steep early drop. Epoch AI's answer is mostly economic rather than technical. Across five benchmarks, the cost of matching a newly state-of-the-art score falls about 66% per quarter at first. Within two years the decline slows to about 32% per quarter. Their explanation is a temporary monopoly. The first lab to reach a capability can charge a premium because nobody else offers it yet. As rivals catch up, that premium disappears, and the early collapse in price is largely the premium being competed away (Do AI performance prices fall faster when capabilities debut?).

The size of the drops is striking. Since 2023, AI performance prices have fallen about 47% per quarter, or roughly 13× per year. That's faster than DNA sequencing costs fell, and far faster than electricity costs ever did. In one example, a benchmark result OpenAI's o3 reached for $0.30 was later matched for $0.0004 (How fast is AI capability becoming cheaper to produce?). A different study helps separate the causes. Researchers looked only at open-weight models, which removes both falling hardware prices and competitive pricing from the picture. What's left, better algorithms on their own, makes inference about 3× more efficient per year (How fast is AI inference actually becoming more efficient?). Compare that 3× with the overall 13× and the rest of the decline has to come from cheaper hardware and from competition shrinking profit margins. The debut period is where that margin shrinkage is most visible.

It helps to know what the price is recovering. Amortized training costs for frontier models have grown about 2.4× per year since 2016 (How fast are frontier AI training costs actually rising?). At data centers, the expensive part is paying off the servers, not the electricity (What actually drives AI data center costs?). The corpus doesn't directly test this next link, but it is a reasonable reading: the short monopoly window after a debut is when labs can recover that heavy spending, which helps explain why launch prices start high.

The less obvious part is that the cost of running the newest model is going up, not down. One analysis finds that performance per dollar improves 5–10× a year, yet the cost of running frontier models rises 3–18× a year. Roughly half of the measured progress on one hard science benchmark comes from spending more on inference rather than from smarter models (Is benchmark progress outpacing the cost to achieve it?). So two curves move at once. Yesterday's best capability gets very cheap very fast, while the state of the art keeps moving to a new, expensive level, and the cycle starts again. A smaller example of the same pattern: compact models trained specifically to carry out long, multi-step work can approach large-model performance at a fraction of the cost (Does model efficiency matter more than peak capability for real work?).

One practical consequence: because prices change this quickly, any dollar cost quoted in a benchmark report goes out of date within months. That's why some researchers argue benchmarks should report how many tokens a task used, so readers can redo the cost math at current prices (Should benchmarks report token counts instead of frozen costs?). There's also a deeper caveat. All of these price curves measure the cost of benchmark scores, and benchmarks reward neat, automatically graded tasks rather than real work (Why do agent benchmarks not predict real economic value?). So the curve shows how fast benchmark capability gets cheaper. It doesn't show how fast useful work gets cheaper.