Epoch AI reports that the cost of achieving a fixed performance benchmark in artificial intelligence has dropped by roughly 13 times annually. This rate of decline surpasses the historical trajectory of any previous technology. MIT researchers attribute about 3x of that improvement to algorithmic progress alone, with the remainder coming from hardware efficiency and market competition. However, current top-tier models are not necessarily cheaper because reasoning tasks require significantly more compute per operation. Selecting a model for practical deployment still depends on quality, speed, and error rates alongside price.
The data suggests that efficiency gains are outpacing raw capability increases in specific cost metrics. This trend lowers the barrier for smaller businesses to access powerful inference without proportional budget increases. Developers must remain vigilant about total compute usage when fine-tuning for complex reasoning tasks.
- 13x annual price drop for fixed benchmarks
- 3x annual contribution from algorithmic progress
- Reasoning models often cost more due to higher compute needs
