Epoch AI estimates that AI performance costs are dropping by about 13 times per year, a pace faster than any previous technology. Even after stripping out hardware improvements and competition, MIT puts annual algorithmic progress at roughly a 3x improvement.

However, the headline drop in cost does not mean every model is cheaper to use. Reasoning models, for instance, can be more expensive because they require far more compute per task. When choosing a model for real-world use, quality, speed, and error rate are just as important as price.