Test-time Compute & Reasoning ☆

Reasoning models produce long internal chains of thought before answering, and accuracy scales with how much compute they spend thinking — a new scaling axis alongside model size and data. o1 introduced it; R1 open-sourced it; every frontier model now has a thinking mode. The infrastructure consequence is enormous: a single hard question can cost thousands of times more tokens than a chat reply, shifting demand from training clusters to inference fleets.