12.1 Reasoning models and test-time compute
You can decide when paying for extra thinking is worth it.
Before:08. Large Language Models
Reasoning models spend extra inference-time compute — longer chains of thought, multiple samples, verification — to buy accuracy on hard problems, and that spend is a product decision with real latency and cost prices attached. It opens the frontier module because test-time compute is the current frontier's defining trade. The default to resist is reasoning mode everywhere: on easy tasks it is slower, costlier and sometimes worse, and the judgement of when to pay is the skill.
Work through these
Chain-of-thought as trained behaviour
Producing intermediate reasoning before answering, learned during training rather than requested in the prompt. This is what distinguishes a reasoning model from a prompt technique.
Test-time scaling and sampling strategies
Spending more computation at answer time rather than at training time, by sampling several attempts or searching among them. It is a different axis of improvement from making the model bigger.
Verifiers, process supervision, self-consistency
Checking candidate answers, rewarding the steps rather than only the result, and taking the consensus of several attempts. These are the mechanisms that make extra computation pay.
Cost and latency trade-offs of reasoning modes
Thinking longer costs more and takes longer, and for many tasks it changes nothing. Deciding where it earns its cost is the practical content of this topic.
Sign in to keep your progress.
Free resources
We haven't checked most of these for screen reader use yet.
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.