Documentation/Core Concepts/Thinking & Reasoning Tokens

Thinking & Reasoning Tokens

Accurately extract and bill hidden reasoning tokens across modern frontier models.


Thinking & Reasoning Tokens

Modern reasoning models (OpenAI o1/o3-mini, GPT-5, Claude 3.7 Thinking, DeepSeek R1) generate hundreds or thousands of internal **reasoning/thinking tokens** that are invisible in the final text response but billed by providers at full output rates.

---

The Hidden Token Problem

If you only count visible output text, a prompt generating 20 words with 4,000 reasoning tokens will cost $0.05 while your meter reports $0.0001—wiping out your profit margins.

`vibezcheck` automatically intercepts and extracts hidden reasoning tokens from provider stream details:

json
{
  "usage": {
    "inputTokens": 650,
    "outputTokens": 2400,
    "totalTokens": 3050,
    "reasoningTokens": 2200,
    "visibleOutputTokens": 200
  },
  "cost": {
    "inputUSD": 0.000715,
    "outputUSD": 0.010560,
    "reasoningCostUSD": 0.009680,
    "totalUSD": 0.011275
  }
}

---

Supported Models * **OpenAI**: `o1`, `o1-mini`, `o3`, `o3-mini`, `gpt-5.6-sol` * **Anthropic**: `claude-3-7-sonnet` (with Extended Thinking) * **DeepSeek**: `deepseek-reasoner` (R1) * **Google**: `gemini-2.0-flash` (Thinking experimental)