Does Claude Extended Thinking Use More Tokens? Cost Explained

Mar 25, 2026

I was confused when my Claude message limit got exhausted after just one complex query. The chat was working fine yesterday, and suddenly I’m hitting rate limits way faster than expected. What happened?

The Problem

The issue was extended thinking (reasoning mode). When extended thinking is enabled, Claude generates internal “thinking” content before producing its final response. These thinking blocks count toward your usage limits just like output tokens.

I discovered this by examining the response structure in my project’s code. In app/services/claude_log_service.py, there’s explicit handling for thinking blocks:

elif block_type == "thinking":
    parsed_blocks.append(
        {
            "event_type": "thinking",
            "title": "Thinking",
            "content": block.get("thinking", ""),
        }
    )

This code confirms that extended thinking generates separate “thinking” blocks that are tracked and stored as additional content, distinct from regular text responses.

Token Consumption Comparison

Let me break down the actual token cost difference:

Standard Response Token Cost:

My input: ~500 tokens
Claude’s response: ~1,000 tokens
System overhead: ~2,000 tokens
Total: ~3,500 tokens

Extended Thinking Response Token Cost:

My input: ~500 tokens
Thinking blocks: ~5,000-15,000 tokens (varies by complexity)
Claude’s response: ~1,000 tokens
System overhead: ~2,000 tokens
Total: ~8,500-18,500 tokens

Extended thinking can multiply token usage by 2-5x per message. No wonder I was hitting limits faster.

Why This Matters

Rate limit exhaustion: Daily/monthly limits deplete 2-5x faster
Cost implications: Paying per token means extended thinking costs more
Feature vs. cost tradeoff: Better reasoning quality comes at a token cost
Model differences: Opus has more powerful thinking than Sonnet but costs more per token

From Anthropic’s pricing page, the model costs are:

Opus 4.6: $5/MTok input, $25/MTok output
Sonnet 4.6: $3/MTok input, $15/MTok output

Common Mistakes I Made

Leaving extended thinking on all the time - I should only enable it when deeper reasoning is needed
Not understanding the token cost - Thinking tokens aren’t free; they count toward limits
Using extended thinking for simple questions - Complete waste for basic queries
Long conversations with thinking enabled - Context accumulates, and thinking compounds the cost

How to Check Token Usage in API Responses

When using the API, you can see the token impact:

# Extended thinking responses include thinking blocks
response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=4096,
    thinking={"type": "enabled", "budget_tokens": 10000},
    messages=[{"role": "user", "content": "Solve this complex problem..."}]
)

# Token usage includes thinking tokens
print(f"Input tokens: {response.usage.input_tokens}")
print(f"Output tokens: {response.usage.output_tokens}")
# Note: thinking tokens are included in output_tokens

To save tokens for simple tasks:

# For simple tasks, disable thinking
response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    thinking={"type": "disabled"},  # Explicit disable
    messages=[{"role": "user", "content": "What is 2+2?"}]
)

Model Selection Strategy

Based on my experience and Anthropic’s pricing, here’s what I recommend:

Scenario	Recommended Model	Extended Thinking?
Simple coding tasks	Sonnet 4.6	Off
Complex debugging	Sonnet 4.6	On
Architecture design	Opus 4.6	On
Quick questions	Sonnet/Haiku	Off
Research and analysis	Opus 4.6	On

Community Insights

I found similar experiences on Reddit. One user noted: “are you keeping thinking or reasoning on at all time? That also has cost. And it eaaaats tokens”

Another mentioned getting “force switched to Sonnet” after hitting limits, which suggests the platform is aware of the token consumption patterns.

Conclusion

Extended thinking in Claude consumes more tokens because it generates additional reasoning content before the final response. These thinking tokens count toward usage limits, causing faster rate limit exhaustion.

To manage costs:

Disable extended thinking for simple queries
Enable it selectively for complex problems
Consider using Sonnet with thinking instead of Opus for cost efficiency
Monitor usage after each message to understand the impact

The tradeoff is real: better reasoning comes at a token cost. Use extended thinking strategically, not by default.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!