The Real Cost of Reasoning: Why Your AI Bill Has Hidden Tokens

You asked GPT-o1 a simple math question. The response was 50 tokens. Your bill shows 2,000 output tokens.

What happened? Reasoning tokens.

The Hidden Token Problem

Modern reasoning models don't just generate answers—they think first. This internal chain-of-thought is invisible to you but very visible on your invoice.

Here's how it works:

Your prompt: "What's 23 × 47?"
                    ↓
    [Internal reasoning - 1,950 tokens]
    "Let me break this down..."
    "23 × 40 = 920"
    "23 × 7 = 161"
    "920 + 161 = 1,081"
    "Let me verify..."
                    ↓
Visible response: "1,081" (50 tokens)
                    ↓
You pay for: 2,000 output tokens

The reasoning trace is hidden. The cost is not.

Which Models Have Hidden Reasoning?

We track reasoning capabilities across 80 providers. Here's the landscape:

Models with hidden reasoning tokens (you pay, you don't see):

Model	Input	Output	Reasoning Cost
OpenAI o1	$15.00	$60.00	Included in output
OpenAI o1-pro	$150.00	$600.00	Included in output
OpenAI o3	$2.00	$8.00	Included in output
OpenAI o3-pro	$20.00	$80.00	Included in output
OpenAI o3-mini	$1.10	$4.40	Included in output

Models with visible reasoning (you pay, you see):

Model	Input	Output	Reasoning
DeepSeek Reasoner	$0.28	$0.42	Visible in response
Claude (extended thinking)	varies	varies	Visible `<thinking>` blocks

Models with no reasoning overhead:

Model	Input	Output
GPT-4o	$2.50	$10.00
GPT-4o Mini	$0.15	$0.60
Claude Sonnet	$3.00	$15.00
Claude Haiku	$1.00	$5.00
Gemini 2.0 Flash	$0.10	$0.40

The Math That Matters

Let's calculate real costs for a reasoning task.

Scenario: You need to solve 100 logic problems. Each problem requires ~2,000 reasoning tokens internally but only ~100 tokens of visible output.

With OpenAI o1:

Input: 500 tokens × 100 = 50,000 tokens → $0.75
Output: 2,100 tokens × 100 = 210,000 tokens → $12.60
Total: $13.35

With OpenAI o3-mini:

Input: 50,000 tokens → $0.055
Output: 210,000 tokens → $0.92
Total: $0.98

With DeepSeek Reasoner:

Input: 50,000 tokens → $0.014
Output: 210,000 tokens → $0.088
Total: $0.10

DeepSeek is 133x cheaper for the same reasoning task.

The "Thinking Budget" Trap

Some models let you set a "thinking budget" or reasoning depth. Higher budgets = more internal tokens = higher costs.

OpenAI's reasoning models have these settings:

reasoning_effort: "low" - Fewer reasoning tokens
reasoning_effort: "medium" - Default
reasoning_effort: "high" - Maximum reasoning

The difference between low and high can be 5-10x in token usage.

Pro tip: Start with low effort

Most tasks don't need maximum reasoning. Test with reasoning_effort: "low" first. Only increase if quality suffers.

When Reasoning Models Make Sense

Worth the cost:

Complex multi-step math
Code debugging with deep logic
Scientific reasoning
Tasks where accuracy > cost

Not worth the cost:

Simple Q&A
Content generation
Summarization
Chat/conversation
Anything a non-reasoning model handles well

The Alternative: Explicit Chain-of-Thought

Instead of paying for hidden reasoning, you can prompt regular models to show their work:

Prompt: "Solve step by step: What's 23 × 47?"

Response:
"Step 1: Break down 47 into 40 + 7
Step 2: 23 × 40 = 920
Step 3: 23 × 7 = 161
Step 4: 920 + 161 = 1,081

Answer: 1,081"

You still pay for the reasoning tokens, but:

You see exactly what you're paying for
You can use cheaper models (GPT-4o Mini at $0.60/1M output vs o1 at $60/1M)
You control the verbosity

Cost comparison for same task:

Approach	Model	Output Cost
Hidden reasoning	o1	$60.00/1M
Hidden reasoning	o3-mini	$4.40/1M
Explicit CoT	GPT-4o	$10.00/1M
Explicit CoT	GPT-4o Mini	$0.60/1M
Visible reasoning	DeepSeek Reasoner	$0.42/1M

Caching Doesn't Help (Much)

Here's the painful truth: reasoning tokens can't be cached.

Caching applies to input tokens—your prompts. The internal reasoning chain is generated fresh each time. Even with identical prompts, the model may reason differently.

This means:

Cached input discount: ✅ Applies
Reasoning token discount: ❌ None

OpenAI's cached input rate is 50% off ($7.50 instead of $15 for o1 input). But output—including all reasoning—stays at full price.

How to Audit Your Reasoning Costs

OpenAI API response structure:

{
  "usage": {
    "prompt_tokens": 500,
    "completion_tokens": 2100,
    "completion_tokens_details": {
      "reasoning_tokens": 2000,
      "accepted_prediction_tokens": 0,
      "rejected_prediction_tokens": 0
    }
  }
}

The reasoning_tokens field tells you exactly how many hidden tokens you're paying for.

Warning signs you're overspending:

reasoning_tokens > 10x your visible output
Simple questions generating thousands of reasoning tokens
Consistent reasoning overhead on repetitive tasks

The Bottom Line

Reasoning models are powerful but expensive. The cost isn't in the visible output—it's in the invisible thinking.

Rules of thumb:

Use reasoning models only when you need reasoning
Start with low effort settings
Monitor reasoning_tokens in your usage data
Consider DeepSeek Reasoner for cost-sensitive reasoning tasks
For simple tasks, explicit chain-of-thought on cheap models beats hidden reasoning on expensive ones

We track pricing for all reasoning models across 80 providers. Compare reasoning model costs →

Pricing data from Subquery. Updated daily across 2,000+ models.