Reasoning Models¶
Some models generate internal "thinking" tokens before answering — working through the problem step by step, then giving you the result. You pay for every thinking token, in money and seconds. Sometimes that's the best money you'll spend; often it's pure waste.
flowchart LR
J["Junior: thinking tokens"] --> M["Middle: when reasoning is waste"]
M --> S["Senior: reasoning vs CoT vs agents"]
S --> P["Professional: governance"]
Levels¶
| Level | Guide | You are done when |
|---|---|---|
| Junior | Thinking tokens | You can explain how a reasoning model works, why you pay for invisible tokens, and what tasks they help. |
| Middle | When reasoning is waste | You can classify your tasks into reasoning-helps vs. reasoning-wastes, and explain the temperature interaction. |
| Senior | Reasoning vs. CoT vs. agents | You can choose between a reasoning model, prompt-level CoT, and an agent loop — and cap the budget. |
| Professional | Governance | You can decide which product surfaces may use reasoning models and measure whether it pays off. |
Practice rule¶
Before reaching for a reasoning model, ask: does this task fail because the model can't figure it out, or because it doesn't have the information? Reasoning helps the first; the second needs retrieval or tools.
Related¶
- How LLMs Work — the loop thinking tokens extend.
- Temperature and Sampling — why reasoning and high temperature fight each other.