GLOSSARY
Test-Time Compute
Spending more compute while answering — thinking longer, sampling many attempts, verifying — instead of only making the model bigger beforehand.
The classic scaling lever was train-time: more parameters, more data. Test-time compute is the newer axis: at question time, a model can deliberate step by step (reasoning models), generate many candidate answers and pick the best (best-of-N), or run verifiers over its own work. o1/R1-class models showed that trading seconds of thinking for accuracy can beat models many times larger.
This reshapes product economics: intelligence becomes rentable per question — hard problems get more compute, easy ones less. Watch for products charging per “thinking level”, and for the flip side: latency and cost per query climb, so routing between fast and deliberate models (as Groq/OpenRouter ecosystems do) becomes a design skill.