Spend the compute
at answer time.

A model's ability used to come almost entirely from training. More data, more parameters, more compute spent once, before anyone touched it. For a decade that was the only dial that mattered, and it was expensive and slow to turn.

Test-time compute is a different dial. It is compute the model spends at inference, after training is finished, on the specific question you asked. Anthropic describes Claude 3.7 Sonnet as both an ordinary model and a reasoning model in one, where thinking is a mode you turn on rather than a separate product. Turn it on and the same model gives itself more time before it answers.

This is a different lever from the two most teams already know. It is not more training and not a bigger prompt. For how a model reads what you give it, see context engineering. Test-time compute changes what the model does after it has read, during the answer itself. A reasoning model is simply a model built to spend this compute well.

Test-time compute comes in two shapes. The first is serial. The model reasons in one long chain, what Anthropic describes as multiple, sequential reasoning steps before producing its final output. Each step builds on the last, and a longer chain is more compute spent on that single line of thought.

The second is parallel. The model runs several independent attempts at once and keeps the best, what Anthropic describes as sampling multiple independent thought processes and selecting the best one. The selection can be a majority vote, a second model acting as a checker, or a learned scoring function that picks the strongest answer without knowing the true one in advance.

Serial goes deeper on one path. Parallel goes wider across many. They are not rivals. Hard problems often use both, a budget of thinking per attempt and a count of attempts to run, and the two together are the compute you are really spending.

A model does not spend test-time compute well by accident. The habit of thinking in steps is trained in, and the clearest public proof came from DeepSeek. In January 2025 it released DeepSeek R1, open under the MIT license, and reported that its R1 Zero variant was trained via large scale reinforcement learning without supervised fine-tuning as a preliminary step.

DeepSeek calls it the first open research to validate that reasoning capabilities can be incentivized purely through reinforcement learning. Left to practice against a reward, the model learned to spend compute on reasoning on its own. DeepSeek reported that R1 reached performance comparable to OpenAI o1 across math, code, and reasoning tasks.

It also shipped the habit at many sizes. DeepSeek open sourced distilled 1.5B, 7B, 8B, 14B, 32B, and 70B checkpoints, so the same reasoning can run on a small, cheap model as well as a large one. Test-time compute is not locked to the frontier. It is a way of spending that smaller models can learn too.