Reasoning Models

Use reasoning models (such as o4-mini, DeepSeek) and control reasoning depth

Reasoning Models

Reasoning Models perform deep thinking before answering, making them suitable for complex tasks like math, coding, and logical reasoning. Crazyrouter supports multiple reasoning models with reasoning depth control.

Supported Reasoning Models

Model	Description
o4-mini	OpenAI reasoning model, balancing speed and capability
o3	OpenAI advanced reasoning model
o3-mini	OpenAI lightweight reasoning model
deepseek-r1	DeepSeek reasoning model
deepseek-v3-1	DeepSeek V3.1
claude-sonnet-4-20250514	Claude with extended thinking support
gemini-2.5-flash-thinking	Gemini thinking model

---

reasoning_effort Parameter

Control the model’s reasoning depth with reasoning_effort:

Value	Description
low	Quick answers, suitable for simple questions
medium	Moderate reasoning, balancing speed and quality
high	Deep reasoning, suitable for complex problems

cURL

curl https://crazyrouter.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "o4-mini",
    "messages": [
      {"role": "user", "content": "Prove that the square root of 2 is irrational"}
    ],
    "reasoning_effort": "high"
  }'

---

thinking Parameter (Thinking Budget)

Some models support precise control of the thinking token budget via the thinking parameter:

Python

response = client.chat.completions.create(
    model="claude-sonnet-4-20250514",
    messages=[
        {"role": "user", "content": "Analyze the time complexity of this algorithm and optimize it"}
    ],
    thinking={
        "type": "enabled",
        "budget_tokens": 10000
    },
    max_tokens=16000
)

When using the thinking parameter, max_tokens must be greater than budget_tokens, because max_tokens includes both thinking tokens and output tokens.

---

Streaming Reasoning

Reasoning models also support streaming output. The thinking process and final answer are returned as separate chunks:

Python

stream = client.chat.completions.create(
    model="o4-mini",
    messages=[
        {"role": "user", "content": "Solve a Sudoku puzzle"}
    ],
    reasoning_effort="medium",
    stream=True
)

for chunk in stream:
    if chunk.choices[0].delta.content is not None:
        print(chunk.choices[0].delta.content, end="")

---

DeepSeek Reasoning Models

DeepSeek reasoning models return the thinking process in the response:

Python

response = client.chat.completions.create(
    model="deepseek-r1",
    messages=[
        {"role": "user", "content": "Which is larger, 9.11 or 9.8?"}
    ]
)

# The thinking process may be in the reasoning_content field
message = response.choices[0].message
print("Answer:", message.content)

Reasoning models typically consume far more tokens than regular models because the thinking process also consumes tokens. Be mindful of controlling reasoning_effort and budget_tokens to manage costs.

Reasoning models typically do not support sampling parameters like temperature and top_p. If these parameters are passed, they may be ignored or return an error.