Adaptive Thinking and Effort: Trading Cost for Answer Quality

Reasoning models can "think" before answering: they work through a problem internally, then write the final response. Thinking improves results on hard tasks such as multi-step maths, debugging and planning, at the cost of extra tokens and time.

Adaptive thinking

With adaptive thinking the model decides for itself how much to think, based on how hard the request is. On current Claude models you control the overall depth with output_config.effort:

const response = await client.messages.create({
  model: 'claude-opus-5-5',
  max_tokens: 16000,
  thinking: { type: 'adaptive', display: 'summarized' }, // show a summary of the reasoning
  output_config: { effort: 'high' }, // low | medium | high | xhigh | max
  messages: [{ role: 'user', content: 'Find the bug in this function and explain it: ...' }],
});

for (const block of response.content) {
  if (block.type === 'thinking') console.log('Reasoning summary:', block.thinking);
  if (block.type === 'text') console.log('Answer:', block.text);
}

Choosing an effort level

• low - chat, classification, extraction, high-volume routes where speed matters.
• medium - a good default for everyday work.
• high / xhigh - coding, agents, long multi-step tasks.
• max - when correctness matters more than cost and you've measured that it helps.

Things to know

• Thinking tokens are billed as output tokens, even when the reasoning text isn't shown to you.
• Set the effort explicitly in code. Defaults differ between models and change between versions.
• Measure on a sample of real requests. Raise effort only on routes where quality clearly improves.
• Use streaming for high-effort requests; long thinking can otherwise look like a hung request.

The old fixed "thinking budget" (budget_tokens) is replaced by adaptive thinking on current models. If you see it in older tutorials, use effort instead.