Once a new AI skill appears, does its price drop on a schedule predictable enough to bet on?

This explores whether AI capabilities get cheaper on a regular schedule after they first appear, and whether that schedule can tell us when they become useful in practice. The short answer is that the price curve has a shape you can predict, but the curve measures less than it seems to. Epoch AI's analysis of five benchmarks found a consistent pattern: the cost of reaching a given level of performance falls about 66% per quarter right after that level becomes state of the art, then slows to about 32% per quarter within two years Do AI performance prices fall faster when capabilities debut?. Their explanation is economic, not technical. A new capability starts out under a temporary monopoly and is priced that way. As rivals catch up, competition wears the premium down. So the schedule is less like Moore's law and more like a drug coming off patent: steep at first, then flattening.

The catch is what's being priced. These curves track the cost of hitting a benchmark score, and benchmark scores turn out to be a poor stand-in for real work. An analysis of 960 real occupational workflows found that agents win abstract contests but fail long, multi-step professional tasks. The authors argue the gap comes from the benchmarks themselves, because the field has measured contests rather than jobs Why do agent benchmarks not predict real economic value?. A smooth price decline on a benchmark can sit right alongside a capability that hasn't matured for actual use. The unit of cost is shifting too. A 115-day case study of a persistent agent found that 82.9% of its tokens were cache reads, which makes cost per token close to meaningless; the useful measure becomes cost per finished piece of work Do persistent agents really cost less per token?. A related line of work shows that compact models trained on execution and follow-through can reach the low-cost sweet spot of the price-performance curve. In other words, efficiency gains can come from training choices rather than from waiting for prices to fall Does model efficiency matter more than peak capability for real work?.

Cheaper capabilities also don't spread evenly. Firms that are more exposed to AI replace freelance labor with AI tools faster and more cheaply than less-exposed firms. That points to returns from building AI capability in-house, not a single shared adoption curve Do firms substitute labor for AI at different rates?. On the economy-wide level, Brynjolfsson, Rock and Syverson's productivity J-curve explains why falling prices may not show up in the numbers for years. Companies first pour money into unmeasured complements like training, process redesign and software. Productivity is undercounted while that happens and overcounted once those investments start paying off Why does AI productivity appear weak when investment is high?. Even a perfectly predictable price curve would show up in measured productivity late, and then all at once.

If you want to see how far 'predictable schedule' thinking can go, look at forecasting models of AI capability. Kwa's 8-parameter model reproduces the aggressive AI Futures Model forecast of near-total automation of AI research by around 2032. It does this by replacing loosely defined assumptions with direct capability measurements Can simpler models predict AI R&D automation timelines accurately?. That's a real case for simple trend-following, and it inherits the same weakness: the trends depend on what gets measured. A more skeptical view argues that as AI output gets cheaper and more plentiful, nothing reliable backs its value. The result is rising quantity alongside falling reliability, so a falling price may partly reflect falling worth What actually backs the value of AI-generated intelligence?.

The surprising part is that the most predictable thing about AI prices is how competition behaves, not how the technology matures. The steep early drop is a monopoly premium disappearing. Whether a capability is actually ready for real work depends on a separate clock, set by how tasks are measured, how much work gets reused, and how long organizations take to build the complements around it. The corpus doesn't yet have direct studies linking price-decay curves to real deployment timelines. That gap is worth noticing.