Anthropic Reasoning

Anthropic Claude models support extended thinking, which lets the model reason through complex problems before producing a final answer. Claude 4.6 models introduce adaptive thinking, where Claude dynamically decides when and how much to think based on an effort level.

These models use thinking: { type: 'adaptive' }. Claude dynamically decides when and how much to think.

These models use thinking: { type: 'enabled', budgetTokens: N } to set a fixed token budget for thinking.

• anthropic/claude-opus-4.5
• anthropic/claude-opus-4.1
• anthropic/claude-opus-4
• anthropic/claude-sonnet-4.5
• anthropic/claude-sonnet-4
• anthropic/claude-haiku-4.5

• Adaptive thinking (Claude 4.6): Use thinking: { type: 'adaptive' }. Claude decides when and how much to think. At high effort (default), Claude almost always thinks. At lower effort levels, it may skip thinking for simpler problems.
• Manual thinking (Claude 4, Opus 4.5): Use thinking: { type: 'enabled', budgetTokens: N } to set a fixed token budget for thinking.

Manual thinking with type: 'enabled' and budgetTokens is deprecated on Claude 4.6 models. It still works but will be removed in a future release. Use adaptive thinking instead.

For more details, see the Anthropic extended thinking docs, adaptive thinking docs, and effort parameter docs.

Configure adaptive thinking through providerOptions. Claude dynamically decides when and how much to think:

import { generateText } from 'ai';
 
const result = await generateText({
  model: 'anthropic/claude-sonnet-4.6',
  prompt: 'Explain quantum entanglement in simple terms.',
  providerOptions: {
    anthropic: {
      thinking: { type: 'adaptive' },
    },
  },
});
 
console.log('Thinking:', result.reasoningText);
console.log('Response:', result.text);

import { streamText } from 'ai';
 
const result = streamText({
  model: 'anthropic/claude-opus-4.6',
  prompt: 'Explain quantum entanglement in simple terms.',
  providerOptions: {
    anthropic: {
      thinking: { type: 'adaptive' },
    },
  },
});
 
for await (const part of result.fullStream) {
  if (part.type === 'reasoning-delta') {
    process.stdout.write(part.text);
  } else if (part.type === 'text-delta') {
    process.stdout.write(part.text);
  }
}

For older models, use type: 'enabled' with a budgetTokens value:

import { generateText } from 'ai';
 
const result = await generateText({
  model: 'anthropic/claude-opus-4',
  prompt: 'Explain quantum entanglement in simple terms.',
  providerOptions: {
    anthropic: {
      thinking: {
        type: 'enabled',
        budgetTokens: 5000,
      },
    },
  },
});
 
console.log('Thinking:', result.reasoningText);
console.log('Response:', result.text);

Interleaved thinking lets Claude think between tool calls, producing better reasoning in multi-step workflows.

• Claude Opus 4.6: Automatically enabled with adaptive thinking
• Claude Sonnet 4.6, Opus 4.5, 4.1, 4, Sonnet 4.5, 4: Pass the interleaved-thinking-2025-05-14 beta header when extended thinking is enabled

import { generateText } from 'ai';
 
const result = await generateText({
  model: 'anthropic/claude-sonnet-4.6',
  prompt: 'Search for the weather and summarize it.',
  providerOptions: {
    anthropic: {
      thinking: { type: 'enabled', budgetTokens: 5000 },
      headers: {
        'anthropic-beta': 'interleaved-thinking-2025-05-14',
      },
    },
  },
  tools: {
    // your tools here
  },
});

With interleaved thinking, budgetTokens can exceed the model's max output tokens since it represents the total budget across all thinking blocks in a single turn.

For more details, see the Anthropic extended thinking docs.

Claude 4 models return summarized thinking output, not full thinking tokens. You're charged for the full thinking tokens, but the response contains a condensed summary. Claude Sonnet 3.7 returns full thinking output.