Why does the price of 'best in class' AI performance crash so much faster than everyday performance improves?

This explores why the price of hitting a top-tier benchmark score falls so steeply, compared with the slower decline for capability levels that are already commonplace. In the corpus, "high-end" is mostly a matter of timing. What matters is how recently a performance level became the best available, not how hard it is in absolute terms. Epoch AI tracked five benchmarks and found that the cost of reaching a given score drops about 66% per quarter while that score is newly state of the art. Within two years, the decline slows to about 32% per quarter Do AI performance prices fall faster when capabilities debut?. Compounded over a year, that is the difference between roughly a 75x price drop and roughly a 5x drop. Today's high-end capability is in its steep phase. Yesterday's high-end, which is today's low-end, has already gone through it.

Epoch's explanation is economic more than technical. When one lab first reaches a capability, nobody else can sell it, so the lab can charge a premium. As competitors catch up, that premium shrinks, and the price falls toward the actual cost of running the model. Most of the early, dramatic drop is that premium disappearing. Once it is gone, prices can only fall as fast as the underlying costs do. The size of the effect shows in the overall figures: AI performance prices have fallen about 47% per quarter since 2023, roughly 13x a year. In one case, a benchmark result that cost $0.30 with o3 was later matched for $0.0004 How fast is AI capability becoming cheaper to produce?.

A useful check is how much of that decline comes from real engineering progress. When researchers remove hardware price declines and competitive pricing effects by looking only at open-weight models, algorithmic efficiency alone improves about 3x per year How fast is AI inference actually becoming more efficient?. That is close to the slow, mature-phase rate. One reading is that 3x a year is roughly the floor every capability eventually settles to, and the much faster early decline for frontier capabilities comes largely from pricing power wearing off. The competition that erodes it is visible too. Open models small enough to run on a consumer GPU gain LM Arena Elo about 56% faster than frontier models Are open models closing on frontier AI faster than expected?. Cheap substitutes keep arriving just behind the frontier, which is what forces new capabilities down in price so quickly.

There is a twist you might not expect. Even as the price of any fixed score collapses, the cost of running the newest frontier models keeps rising, 3x to 18x a year. About half of the measured progress on GPQA-Diamond tracks more inference spending rather than smarter models Is benchmark progress outpacing the cost to achieve it?. So the high end is a moving target. Each new top score often begins at a high price because more compute was spent to reach it, which leaves room for the steep decline that follows. That is also why frozen dollar figures in benchmark reports go stale quickly, and why researchers argue that reports should include token counts so costs can be recalculated as prices fall Should benchmarks report token counts instead of frozen costs?. A related problem is that leaderboards that score only accuracy push toward needlessly expensive systems Why do accuracy-only benchmarks hide the true cost of agents?.

One limit on what the corpus can answer: it explains the pattern mainly through how long a capability has been state of the art, with a temporary monopoly that fades. It doesn't directly test whether harder benchmarks are inherently cheaper to improve. If your question is about difficulty rather than timing, the material here is thinner.