In the "old days" of 2024, billing for Large Language Models (LLMs) was simple: you paid for what you sent (Input) and what you got back (Output). In 2025, the rise of Reasoning Models (like OpenAI o1, DeepSeek-R1, and Google Gemini 2.0 Thinking) has introduced a third, often invisible billing dimension: Reasoning Tokens.

This article deconstructs the new anatomy of an inference call and explains why your "short answer" might be costing you a fortune.

The New Billing Equation

Traditional LLM billing looked like this:

Cost=(Input Tokens×P in ​ )+(Output Tokens×P out ​ )

The 2025 equation for reasoning models adds a critical component:

Cost=(Input×P in ​ )+(Visible Output×P out ​ )+(Reasoning Tokens×P reasoning ​ )

The "Phantom" Tokens

Reasoning tokens are generated by the model to "think through" a problem using Chain-of-Thought (CoT) logic. Crucially, these tokens are often discarded before the final response is sent to the user, but they are billed as output tokens.

Case Study: The "Simple" Math Problem

• User Prompt: "Solve this differential equation..." (50 tokens)

• Model Process:

  • Step 1 (Internal): Decomposes equation. (500 tokens)

  • Step 2 (Internal): Attempts method A, fails. (1,000 tokens)

  • Step 3 (Internal): Backtracks, tries method B, succeeds. (2,000 tokens)

• Final Answer: "The solution is $x=5$." (10 tokens)

The Billable Event: You see 60 tokens (50 in + 10 out). You are billed for 3,560 tokens.

Managing Reasoning Depth

API providers have introduced parameters to control this hidden cost.

1. reasoning_effort (OpenAI/Azure):

   • low: Forces the model to take fewer logical steps. Good for quick validations.

   • high: Allows deep exploration. Essential for coding or scientific proofs but can spike costs by 10x.

2. max_reasoning_tokens (DeepSeek/Anthropic): A hard cap on the internal scratchpad. If the model hits this limit without finding an answer, it returns its best guess or an error.

Architectural Best Practice: The "Verifier" Pattern

To optimize costs, avoid asking the reasoning model to generate simple text. Use it only to verify or structure complex logic.

• Expensive: Ask o1 to "Write a blog post about quantum physics." (It will overthink the structure, tone, and facts).

• Efficient: Ask o1 to "Create a detailed outline and fact-list for a blog post about quantum physics." Then, pass that outline to a cheaper model (GPT-4o or Llama 3.2) to do the actual writing.

Conclusion: In 2025, engineering leaders must audit their logs not just for response length, but for total generation length. If you aren't tracking reasoning tokens, your unit economics are wrong.