Test-Time Compute
Also called: Inference-Time Scaling
Test-time compute is the practice of allocating extra computation at the moment a model answers rather than during training, letting it explore several reasoning paths or search trees before committing to a response. It shifts investment from very large pre-training runs to targeted reasoning at query time, which allows a smaller model to reach accuracy comparable to a much larger one on logical, mathematical and coding problems.
In practice
This changes where the money sits: less capital in training, more variable cost per query, and a bill that scales with usage rather than with headcount. Model the runtime cost per query before rollout, because a workload that looked affordable in a pilot can move the unit economics once every user is asking hard questions.