Reasoning Models: When a Model Thinks Before It Answers

Some models pause and work through a problem step by step before answering - visibly or invisibly. That extra thinking costs time and tokens, and it doesn't always pay off.

What a reasoning model does differently

A standard model generates its answer directly, token by token, from your prompt. A reasoning model - like Claude with extended thinking switched on, OpenAI's gpt-5.6 in a higher reasoning-effort mode, or DeepSeek's R1 line - first generates an internal chain of reasoning, then produces the final answer informed by that reasoning. Depending on the product, you can see that reasoning trace, or it stays hidden.

The thinking budget

Most reasoning models let you control how much thinking happens - a token budget, or an effort level like low/medium/high. A bigger budget means more room to work through a hard problem, but it also means more tokens spent and more time before the first visible output.

Where it actually pays off

Planning a multi-step task, debugging a subtle logic error, working through math or a puzzle with several dependent steps - reasoning tends to help most where a wrong first instinct is common and a bit of deliberate checking catches it.

Where it's often wasted

A simple edit, a short factual lookup, a one-line fix - reasoning mode on a task like that just burns tokens and adds latency for basically the same answer you'd have gotten directly. Good tooling routes trivial requests to standard mode and saves reasoning for tasks that actually need it.

EXAMPLE

Claude API call with a thinking budget: { "model": "claude-sonnet-5", "thinking": {"type": "enabled", "budget_tokens": 4000}, "messages": [{"role": "user", "content": "Find the bug in this recursive function and explain why it fails on n=0."}] } Response includes a thinking block reasoning through base cases before the final answer.

🛠️ EXERCISE — TRY IT YOURSELF

Compare a standard response against a reasoning-mode response on both an easy and a hard task.

Pick one trivial task (e.g. "capitalize this sentence") and one genuinely tricky one (e.g. a multi-step logic puzzle or a subtle bug).
Run both tasks with reasoning/extended thinking turned off, and note the time and quality.
Run both tasks again with reasoning turned on, using a moderate thinking budget or effort level.
Compare token usage and response time for each task, with and without reasoning.
Judge: did reasoning meaningfully improve the hard task? Did it change anything on the easy one besides speed?

✅ SELF-CHECK

☐ Did reasoning mode measurably improve the answer to the hard task compared to the standard pass?
☐ Did reasoning mode noticeably increase latency or tokens on the trivial task without improving the result?

QUICK QUIZ

What's the main tradeoff when you turn on a reasoning/thinking mode for a simple task?