跳转到主要内容

​
AI Thinking Fields

This page only lists thinking and reasoning surfaces that were revalidated against Crazyrouter production on 2026-03-22. The key point is that different protocols expose “thinking” differently:
  • OpenAI Responses returns a standalone reasoning item
  • Claude native Messages returns a thinking block
  • Gemini Native is most reliably observed through usageMetadata.thoughtsTokenCount

​
Currently verified fields

ModelProtocolVerified field or markerNotes
gpt-5.4Responsesoutput[].type = "reasoning"Can return displayable summaries when summary is requested
claude-opus-4-6-thinkingAnthropic Messagescontent[].type = "thinking"Returned alongside a text block
gemini-3-pro-previewGemini NativeusageMetadata.thoughtsTokenCountConfirms thinking tokens were actually used
In the current recheck, gpt-5.4 Chat Completions did not reliably return usable message.reasoning_content, so it is not treated as the primary observable field here.

​
GPT: Responses reasoning item

{
  "output": [
    {
      "id": "rs_xxx",
      "type": "reasoning",
      "encrypted_content": "...",
      "summary": [
        {
          "type": "summary_text",
          "text": "..."
        }
      ]
    },
    {
      "type": "message",
      "content": [
        {
          "type": "output_text",
          "text": "Final answer"
        }
      ]
    }
  ]
}
Extraction example: 
response = client.responses.create(
    model="gpt-5.4",
    input="Which is larger, 9.11 or 9.9?",
    reasoning={"effort": "high", "summary": "detailed"}
)

for item in response.output:
    if item.type == "reasoning":
        for part in item.summary:
            if part.type == "summary_text":
                print("Thinking summary:", part.text)

​
Claude: thinking block

{
  "content": [
    {
      "type": "thinking",
      "thinking": "..."
    },
    {
      "type": "text",
      "text": "Final answer"
    }
  ]
}
Extraction example: 
message = client.messages.create(
    model="claude-opus-4-6-thinking",
    max_tokens=320,
    thinking={
        "type": "enabled",
        "budget_tokens": 128
    },
    messages=[
        {"role": "user", "content": "Which is larger, 9.11 or 9.9?"}
    ]
)

for block in message.content:
    if block.type == "thinking":
        print("Thinking:", block.thinking)
    elif block.type == "text":
        print("Answer:", block.text)

​
Gemini: thoughtsTokenCount

In the current production recheck, the most reliable observable signal was not a visible reasoning paragraph in the body, but a usage field: 
{
  "usageMetadata": {
    "thoughtsTokenCount": 120
  }
}
This indicates that:
  • the thinking budget was actually used
  • the request consumed thinking tokens during generation
Extraction example: 
data = response.json()
thoughts = data.get("usageMetadata", {}).get("thoughtsTokenCount", 0)
print("Thinking tokens:", thoughts)

​
Usage notes

  • Do not assume every model returns raw chain-of-thought text
  • Do not mix thinking fields across protocols
  • For logging, auditing, or UI display, decide per protocol whether to read reasoning, thinking, or thoughtsTokenCount
Related pages: