Test-time compute is the idea that an AI answers better if you let it think longer before speaking. Instead of blurting an instant answer, the model "thinks" for several seconds or minutes — it explores several paths, verifies its calculations, backtracks when needed. OpenAI popularized the approach with o1 in 2024, then every major lab adopted it. It costs more tokens but results are noticeably better on math, code and complex reasoning.