Loading page content
Research
AI cost decline outpaces earlier technologies, but today's top models aren't necessarily cheaper
The Decoder reports Epoch AI's estimate that AI prices fall about 13x per year once a fixed benchmark level is reached, citing about 3x annual algorithmic progress per MIT.
What the numbers say
The Decoder reports that AI performance costs are falling faster than those of any previous technology, citing Epoch AI's measure of about a 13x-per-year price decline as AI hits a fixed benchmark performance level. After stripping out hardware gains and competition, MIT puts annual algorithmic progress at about 3x, according to the piece. It cautions that the trend does not mean today's best models are cheaper: reasoning models can cost more because they use far more compute per task. [1]
Sources
- AI performance costs are falling faster than those of any previous technology
The Decoder · Reporting ·
Related stories
All storiesAnthropic says Claude found a new enzyme system in bacteriophage DNA
Anthropic says Claude discovered a previously unknown enzyme system in bacteriophage DNA, next to a CRISPR-like repeating array whose function is unknown.
Study: AI experts lowballed progress speed
The Decoder reports a Forecasting Research Institute study found AI experts consistently underestimated progress, with IMO gold reached five years early.
arXiv paper: ask coding agents if a stored claim still holds, not if a diff preserves behavior
An arXiv preprint reports that judging a specific stored claim beats judging behavior-preserving diffs when invalidation precision rises from 0.29-0.33 to 0.71-0.97.