Test-time scaling means making an AI smarter at inference time — not by retraining it, but by giving it more compute to think during a response. Instead of one quick pass, the model reasons through multiple paths, checks its work, or generates many candidate answers and picks the best.

This matters because training large models is extraordinarily expensive and slow. Test-time scaling offers a different lever: spend more compute per question to get better answers from the same model. It's why models like OpenAI's o-series feel noticeably stronger on hard reasoning tasks — they're essentially thinking longer before responding.

The tradeoff is cost and latency. A model that "thinks harder" uses more resources per query, making it slower and pricier. Researchers are working on efficiency techniques like KV-cache transfer to reduce redundant computation across reasoning steps.