Compute2025 · Compute · 2 min read · Epoch AI

Inference cost per token keeps falling

Prices for a fixed capability level drop 9×–900× per year — on-premise deployment for public institutions gets cheaper every quarter.

Any product that needs to speak Karakalpak can now call an API instead of building a voice from nothing.

The model works. It does not work well enough for the things people would immediately use it for, so it stays in research.

Debt notifications are now sent automatically, through the right channel, with a record of what was sent and when.