The idea
Reasoning models are trained, usually with reinforcement learning, to produce long internal reasoning before the final answer. They do chain of thought by themselves, check their work and try alternatives. You typically control how much they think with a budget or effort setting.
They shine on maths, complex code, planning and analysis with many constraints. They are slower to first token, use many more tokens and cost more. For chat, extraction and simple classification, a fast standard model is usually better value.
Prompt them differently: state the goal and constraints clearly, and skip the "think step by step" scaffolding they already do. Give them enough output budget so reasoning does not crowd out the answer.
Picking reasoning only where it pays off is one of the biggest cost and latency decisions in an AI product.
Words to know
- Thinking budgetMax tokens the model may spend reasoning.
- Test-time computeSpending more computation while answering, rather than in training.
Try it without code
Hard vs easy task
- In an app that lets you toggle extended thinking, ask a simple factual question with thinking on and off.
- Then ask a tricky scheduling puzzle both ways.
- Note speed and quality for each.
Use it in code
Turn on extended thinking with a budget and read the separate blocks.
import anthropic
client = anthropic.Anthropic()
msg = client.messages.create(
model="claude-sonnet-5-5",
max_tokens=8000,
thinking={"type": "enabled", "budget_tokens": 4000},
messages=[{"role": "user", "content": "Schedule 5 meetings (A-E) in 3 rooms over 2 hours. A and B share a person; C must follow A; D needs room 1. Give one valid plan."}],
)
for block in msg.content:
print(block.type, "->", (getattr(block, "thinking", None) or getattr(block, "text", ""))[:300])
print(msg.usage)Run 5 hard and 5 easy prompts on a reasoning and a standard model. Record accuracy, latency and cost in a table.
Common mistakes
- Defaulting every request to a reasoning model.
- Setting max_tokens too low, so thinking eats the whole budget.