AI models keep getting cheaper to run — but do falling prices actually track real efficiency gains, or just competition and cheaper chips?

This explores two ways of measuring how fast AI gets cheaper to run. One tracks how much energy a model needs for a task. The other works backward from what providers charge. The short answer from the corpus is that the two methods agree surprisingly well, at roughly 3× improvement per year. But that only holds after a lot of cleanup on the price side. The corpus has one direct source on this comparison, and the more interesting material is about why prices are such a noisy signal to begin with.

The main finding comes from work estimating algorithmic progress in inference How fast is AI inference actually becoming more efficient?. Raw API prices fall for three tangled reasons: chips get cheaper, providers undercut each other, and the algorithms themselves get better. To isolate the algorithmic part, the researchers looked at open-weight models, which anyone can run, so a vendor's pricing strategy can't hide their true cost. They then removed the effect of falling hardware prices. What remained was about 3× per year, close to what independent energy-efficiency measurements find. Energy measures physical work rather than a market price, so it can't be skewed by price wars or discounts. When two methods with completely different weak spots land on the same number, you can trust that number more.

Why are prices so far removed from the underlying physics? One reason is that energy is a small part of what AI actually costs. Epoch AI's model of a one-gigawatt data center finds servers make up about 60% of annual cost, while electricity is only about 7% What actually drives AI data center costs?. So a price mostly reflects paying off hardware, plus profit margin and competition. Energy is a real but minor ingredient. That means energy efficiency and price efficiency could easily drift apart. A model could halve its energy use and barely move its price, or drop in price only because a competitor arrived. The fact that they line up anyway is notable.

The picture gets messier when you measure price per benchmark point instead of price per token. Price-performance on benchmarks improves 5–10× per year, yet the total cost of running frontier models rises 3–18× per year. Roughly half of the measured gains on one hard science benchmark come from spending more compute at inference time, not from smarter models Is benchmark progress outpacing the cost to achieve it?. So "efficiency" can mean cheaper per unit of work, or it can mean more capability bought with more money. Price-based estimates can blur those two together unless they are carefully separated.

The choice of unit you divide by also matters a lot. In a 115-day study of a long-running agent, 82.9% of tokens were reads from a cache, so cost per token stopped meaning much. Cost per finished piece of work became the more useful measure Do persistent agents really cost less per token?. Agent benchmarks that score only accuracy hide cost altogether, which pushes developers toward needlessly expensive designs Why do accuracy-only benchmarks hide the true cost of agents?. On phones, the main cost is often moving model weights through memory rather than doing the math. Re-running the same block of the model twice can be cheaper than loading new weights Does recomputing weights cost less than moving them on mobile?. Neither a price sheet nor a simple energy-per-token figure would capture that. Agreeing on "3× per year" is real progress, but whether it describes your situation depends on what you count: tokens, benchmark points, finished work, or bytes moved through memory.