Module 3 · Running models in production · Lesson 8
Reasoning models and extended thinking
Why letting a model generate tokens before it answers makes it smarter, how reinforcement learning taught models to think at length, and how to control, pay for and trust extended thinking in Claude.
By the end of this lesson you can
- Explain why intermediate tokens add computation a single forward pass cannot provide
- Describe test-time compute scaling (chain of thought, self-consistency, learned long reasoning) and the evidence for it
- Explain how RL with verifiable rewards produced reasoning models such as DeepSeek-R1
- Configure Claude's adaptive thinking and effort, and account for thinking tokens in cost, latency and caching
- Decide from marginal-cost arithmetic when a higher effort level is worth paying for