Fix Claude Extended Thinking Not (2026)

Extended thinking gives Claude deeper reasoning capabilities, but misconfigured parameters produce 400 errors or empty thinking blocks. This guide covers every failure mode and the exact fix.

The Error

{
 "type": "error",
 "error": {
 "type": "invalid_request_error",
 "message": "thinking.budget_tokens: must be >= 1024 and < max_tokens"
 }
}

Quick Fix

  1. Set budget_tokens to at least 1024 and strictly less than max_tokens.
  2. When using tools with thinking, set tool_choice to auto or none only.
  3. Pass thinking blocks back unmodified in multi-turn conversations.

What Causes This

Extended thinking fails when:

  • budget_tokens is less than 1024 or greater than or equal to max_tokens.
  • tool_choice is set to any or a specific tool name (only auto and none work with thinking).
  • Thinking is toggled on or off mid-assistant-turn.
  • Thinking blocks are modified or stripped when passing them back in multi-turn conversations.
  • Thinking parameters change between turns, invalidating cached messages.

Full Solution

Basic Extended Thinking Setup

import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000},
 messages=[{"role": "user", "content": "Solve this step by step: What is 127 * 389?"}]
)
for block in response.content:
 if block.type == "thinking":
 print(f"Thinking: {block.thinking[:200]}...")
 elif block.type == "text":
 print(f"Answer: {block.text}")

Fix budget_tokens Validation

The budget_tokens value must satisfy: 1024 <= budget_tokens < max_tokens:

# WRONG: budget_tokens >= max_tokens
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=8000,
 thinking={"type": "enabled", "budget_tokens": 8000}, # Error: not < max_tokens
 messages=[...]
)
# WRONG: budget_tokens < 1024
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=8000,
 thinking={"type": "enabled", "budget_tokens": 500}, # Error: < 1024
 messages=[...]
)
# CORRECT
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000}, # 1024 <= 10000 < 16000
 messages=[...]
)

Fix Tool Choice Conflicts

Extended thinking only supports tool_choice: auto or tool_choice: none:

# WRONG: tool_choice "any" with thinking
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000},
 tool_choice={"type": "any"}, # Error!
 tools=[{"name": "calc", "description": "Calculate", "input_schema": {"type": "object", "properties": {}}}],
 messages=[...]
)
# CORRECT: tool_choice "auto" with thinking
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000},
 tool_choice={"type": "auto"}, # OK
 tools=[{"name": "calc", "description": "Calculate", "input_schema": {"type": "object", "properties": {}}}],
 messages=[...]
)

Control Thinking Display

By default, Claude 4 models return summarized thinking. You can control this with the display parameter:

# Summarized thinking (default) -- charged for full tokens, returns summary
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000, "display": "summarized"},
 messages=[...]
)
# Omit thinking content -- returns empty thinking blocks with encrypted signature
response = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000, "display": "omitted"},
 messages=[...]
)

Multi-Turn Thinking Continuity

Pass thinking blocks back unmodified to maintain reasoning continuity:

# First turn
response1 = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000},
 messages=[{"role": "user", "content": "What is 127 * 389?"}]
)
# Second turn -- pass ALL content blocks back unmodified
messages = [
 {"role": "user", "content": "What is 127 * 389?"},
 {"role": "assistant", "content": response1.content}, # Includes thinking blocks
 {"role": "user", "content": "Now multiply that result by 2"}
]
response2 = client.messages.create(
 model="claude-sonnet-4-6",
 max_tokens=16000,
 thinking={"type": "enabled", "budget_tokens": 10000},
 messages=messages
)

TypeScript Example

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic();
const response = await client.messages.create({
 model: "claude-sonnet-4-6",
 max_tokens: 16000,
 thinking: { type: "enabled", budget_tokens: 10000 },
 messages: [{ role: "user", content: "Solve step by step: What is 127 * 389?" }]
});
for (const block of response.content) {
 if (block.type === "thinking") {
 console.log("Thinking:", block.thinking.slice(0, 200));
 } else if (block.type === "text") {
 console.log("Answer:", block.text);
 }
}

Prevention

  1. Always set max_tokens > budget_tokens + expected output: A good rule is max_tokens = budget_tokens + 4096.
  2. Default to tool_choice auto: When combining tools with thinking, always use auto.
  3. Never modify thinking blocks: In multi-turn conversations, return them exactly as received.
  4. Keep thinking params stable: Changing thinking parameters between turns invalidates cached messages but not cached system prompts or tools.
I hit this exact error six months ago. Then I wrote a CLAUDE.md that tells Claude my stack, my conventions, and my error handling patterns. Haven't seen it since. I run 5 Claude Max subs, 16 Chrome extensions serving 50K users, and bill $500K+ on Upwork. These CLAUDE.md templates are what I actually use. Not theory — production configs. **[Grab the templates — $99 once, free forever →](https://zovo.one/lifetime?utm_source=ccg&utm_medium=cta-error&utm_campaign=claude-extended-thinking-not-working)** 47/500 founding spots. Price goes up when they're gone.

Frequently Asked Questions

What models support extended thinking?

Extended thinking is available on Claude Opus 4.6 and Claude Sonnet 4.6. It is not available on Claude Haiku or older model versions. When using Claude Code, the model is selected either by default or through the --model flag. Verify your model supports thinking before enabling the parameter.

Does extended thinking cost more than regular responses?

Yes. Thinking tokens are billed as output tokens even though they represent internal reasoning rather than the final answer. A request with 10,000 thinking tokens and 2,000 output tokens is billed for 12,000 output tokens total. However, tasks that require multiple retries without thinking often cost more in aggregate than a single successful thinking-enabled attempt.

Can I see what Claude is thinking?

By default, Claude 4 models return summarized thinking. You can control this with the display parameter: "summarized" returns a brief summary of the reasoning, "omitted" returns empty thinking blocks with an encrypted signature for multi-turn continuity. The full internal reasoning is not exposed to prevent prompt injection through reasoning manipulation.

Why does thinking fail silently without producing better results?

If thinking is enabled but the budget is set too low (close to 1024), Claude may not have enough tokens to reason deeply and falls back to a fast response. Set the thinking budget to at least 5,000 tokens for meaningful reasoning improvement. For complex multi-step problems, use 10,000 or more.

Which model? → Take the 5-question quiz in our Model Selector.

Try it: Paste your error into our Error Diagnostic for an instant fix.