Reasoning Models

ZenMux provides fine-grained control over model reasoning behavior and supports multiple API protocols:

OpenAI Chat Completion API: use the reasoning_effort or reasoning parameter
OpenAI Responses API: use the reasoning parameter to control reasoning intensity and summaries
Anthropic Messages API: use the thinking parameter to enable Extended Thinking
Google Vertex AI API: use the thinking_config parameter to control Thinking Mode

With these parameters, you can flexibly tune the model’s reasoning depth and resource allocation based on task complexity.

OpenAI Chat Completion API

Parameter Reference

reasoning_effort

A reasoning-intensity parameter that follows the OpenAI protocol, used to control the depth of the model’s reasoning.

Available values:

low: low-intensity reasoning, suitable for simple tasks
medium: medium-intensity reasoning (default)
high: high-intensity reasoning, suitable for complex tasks
minimal: minimal reasoning intensity

Default behavior

If you don’t pass this parameter, the system uses medium by default.

reasoning

The reasoning parameter provides more fine-grained reasoning control and supports the following fields:

json

{
  "reasoning": {
    "effort": "medium",
    "max_tokens": 1024,
    "enabled": true
  }
}

effort

Equivalent to the reasoning_effort parameter, intended for models that only support reasoning_effort. Available values are the same as reasoning_effort.

max_tokens

Limits the maximum number of reasoning tokens, intended for models that support a thinking budget. This lets you precisely control how many tokens the model can spend during the reasoning phase.

enabled

Controls whether reasoning is enabled. The default is true; set to false to disable reasoning.

Parameter Priority and Auto-Calculation

ZenMux automatically computes and fills in the parameters required by the model based on what you provide, ensuring optimal reasoning behavior.

Default parameter completion

When neither reasoning_effort nor reasoning is provided, the system automatically applies the following default configuration:

json

{
  "reasoning_effort": "medium",
  "reasoning": {
    "effort": "medium"
  }
}

Automatic max_tokens calculation

When the user specifies max_completion_tokens, or when the model itself has a max_completion_tokens limit, the system automatically calculates reasoning.max_tokens based on reasoning.effort.

Rules:

low:    20% of max_completion_tokens
medium: 50% of max_completion_tokens
high:   80% of max_completion_tokens

Reverse inference of effort

When the user provides max_tokens but does not specify effort, the system infers effort using the following rules:

Compute the ratio: reasoning.max_tokens / max_completion_tokens
Compare the ratio against the standard tiers (20%, 50%, 80%)
Choose the closest tier as the effort value

Example: If reasoning.max_tokens / max_completion_tokens = 30%, the system will automatically set effort to low.

API Call Examples

Native call with the OpenAI Python SDK

Use the native OpenAI Python SDK call pattern and pass the reasoning_effort parameter directly:

python

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://zenmux.ai/api/v1",
    api_key=os.getenv("ZENMUX_API_KEY"),
)

completion = client.chat.completions.create(
    model="qwen/qwen3-max-preview",
    reasoning_effort="high",
    messages=[
        {
          "role": "user",
          "content": "What is the meaning of life?"
        }
    ]
)
print(completion.choices[0])

Limitations of the native call pattern

When using the OpenAI SDK’s native reasoning_effort parameter, you cannot precisely control the number of reasoning tokens. This approach only supports reasoning intensity levels (low/medium/high) and does not support precise max_tokens configuration. For more fine-grained control, use the advanced approach.

Advanced usage with the OpenAI Python SDK

Achieve finer-grained reasoning control via the extra_body parameter:

python

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://zenmux.ai/api/v1",
    api_key=os.getenv("ZENMUX_API_KEY"),
)

response = client.chat.completions.create(
    model="qwen/qwen3-max-preview",
    messages=[
        {"role": "user", "content": "How would you build the world's tallest skyscraper?"}
    ],
    extra_body={
        "reasoning": {
            "effort": "high",
            "max_tokens": 2048
        }
    },
)

msg = response.choices[0].message
print(getattr(msg, "reasoning", None))
print(msg.content)

cURL

Using the reasoning parameter

bash

curl https://zenmux.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $ZENMUX_API_KEY" \
  -d '{
    "model": "qwen/qwen3-max-preview",
    "messages": [
      {
        "role": "user",
        "content": "What is the meaning of life?"
      }
    ],
    "reasoning": {
      "effort": "high",
      "max_tokens": 1024
    }
  }'

Using the reasoning_effort parameter

bash

curl https://zenmux.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $ZENMUX_API_KEY" \
  -d '{
    "model": "qwen/qwen3-max-preview",
    "messages": [
      {
        "role": "user",
        "content": "What is the meaning of life?"
      }
    ],
    "reasoning_effort": "high"
  }'

Extracting reasoning output

Retrieve the model’s reasoning process to understand how it arrived at the final answer:

python

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://zenmux.ai/api/v1",
    api_key=os.getenv("ZENMUX_API_KEY"),
)

response = client.chat.completions.create(
    model="qwen/qwen3-max-preview",
    messages=[
        {"role": "user", "content": "Solve this math problem: 2x + 5 = 15"}
    ],
    reasoning_effort="high",
)

message = response.choices[0].message

# Check whether reasoning content exists
if hasattr(message, 'reasoning') and message.reasoning:
    print("Reasoning process:")
    print(message.reasoning)
    print("\nFinal answer:")
    print(message.content)
else:
    print("Direct answer:")
    print(message.content)

Best practices

For simple factual questions, low or medium reasoning effort is usually sufficient
For tasks that require complex logical deduction (e.g., math problems, code generation), high reasoning effort is recommended
Use the max_tokens parameter to control cost while maintaining reasoning quality

OpenAI Responses API

The OpenAI Responses API uses the reasoning parameter to control reasoning behavior. It is similar to the Chat Completion API, but the parameter placement differs slightly.

Parameter Reference

reasoning

json

{
  "reasoning": {
    "effort": "medium",
    "summary": "auto"
  }
}

effort: controls reasoning intensity; supports none / minimal / low / medium / high / xhigh
summary: reasoning summary level; supports auto / concise / detailed

API Call Examples

PythonTypeScriptcURL

python

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://zenmux.ai/api/v1",
    api_key=os.getenv("ZENMUX_API_KEY"),
)

response = client.responses.create(
    model="openai/gpt-5", 
    input="Solve this math problem: If x^2 + 5x + 6 = 0, what are the values of x?",
    reasoning={ 
        "effort": "high", 
        "summary": "detailed"
    }
)

# Get reasoning and output
for item in response.output:
    if item.type == "reasoning":
        print("Reasoning summary:", item.summary)
    elif item.type == "message":
        for content in item.content:
            if content.type == "output_text":
                print("Final answer:", content.text)

import OpenAI from "openai";

const openai = new OpenAI({
  baseURL: "https://zenmux.ai/api/v1",
  apiKey: process.env.ZENMUX_API_KEY,
});

async function main() {
  const response = await openai.responses.create({
    model: "openai/gpt-5", 
    input:
      "Solve this math problem: If x^2 + 5x + 6 = 0, what are the values of x?",
    reasoning: {
      effort: "high", 
      summary: "detailed",
    },
  });

  // Get reasoning and output
  for (const item of response.output) {
    if (item.type === "reasoning") {
      console.log("Reasoning summary:", item.summary);
    } else if (item.type === "message") {
      for (const content of item.content) {
        if (content.type === "output_text") {
          console.log("Final answer:", content.text);
        }
      }
    }
  }
}

main();

bash

curl https://zenmux.ai/api/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $ZENMUX_API_KEY" \
  -d '{
    "model": "openai/gpt-5",
    "input": "Solve this math problem: If x^2 + 5x + 6 = 0, what are the values of x?",
    "reasoning": {
      "effort": "high",
      "summary": "detailed"
    }
  }'

Response Format

The Responses API returns reasoning content blocks in the output array:

json

{
  "output": [
    {
      "type": "reasoning",
      "summary": ["Factor the quadratic...", "Use factoring..."]
    },
    {
      "type": "message",
      "content": [
        {
          "type": "output_text",
          "text": "x = -2 or x = -3"
        }
      ]
    }
  ]
}

Anthropic Messages API (Extended Thinking)

Anthropic Claude models provide reasoning capabilities via Extended Thinking. Enable it with the thinking parameter; the model will perform deep reasoning before answering.

Supported Models

Claude Sonnet 4.5
Claude Sonnet 4
Claude Haiku 4.5
Claude Opus 4.5 / 4.1 / 4

Parameter Reference

thinking

json

{
  "thinking": {
    "type": "enabled",
    "budget_tokens": 10000
  }
}

type: set to "enabled" to enable Extended Thinking
budget_tokens: the thinking token budget, which controls the maximum number of tokens the model can use for reasoning

Notes

budget_tokens must be less than max_tokens
When thinking is enabled, modifying temperature or top_k is not supported
Response pre-fill is not supported
Forced tool calling is not supported (tool_choice: "any" or a specific tool name)

API Call Examples

PythonTypeScriptcURL

python

import anthropic

client = anthropic.Anthropic(
    api_key="<your ZENMUX_API_KEY>", 
    base_url="https://zenmux.ai/api/anthropic"
)

message = client.messages.create(
    model="anthropic/claude-sonnet-4.5", 
    max_tokens=16000,
    thinking={ 
        "type": "enabled", 
        "budget_tokens": 10000
    },
    messages=[
        {
            "role": "user",
            "content": "Are there an infinite number of prime numbers such that n mod 4 == 3?"
        }
    ]
)

# Extract the thinking process and the final answer
for block in message.content:
    if block.type == "thinking":
        print("Thinking process:", block.thinking[:500], "...")
    elif block.type == "text":
        print("Final answer:", block.text)

import Anthropic from "@anthropic-ai/sdk";

const anthropic = new Anthropic({
  apiKey: "<your ZENMUX_API_KEY>", 
  baseURL: "https://zenmux.ai/api/anthropic", 
});

async function main() {
  const message = await anthropic.messages.create({
    model: "anthropic/claude-sonnet-4.5", 
    max_tokens: 16000,
    thinking: {
      type: "enabled", 
      budget_tokens: 10000, 
    },
    messages: [
      {
        role: "user",
        content:
          "Are there an infinite number of prime numbers such that n mod 4 == 3?",
      },
    ],
  });

  // Extract the thinking process and the final answer
  for (const block of message.content) {
    if (block.type === "thinking") {
      console.log("Thinking process:", block.thinking.slice(0, 500), "...");
    } else if (block.type === "text") {
      console.log("Final answer:", block.text);
    }
  }
}

main();

bash

curl https://zenmux.ai/api/anthropic/v1/messages \
  -H "x-api-key: $ZENMUX_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-sonnet-4.5",
    "max_tokens": 16000,
    "thinking": {
      "type": "enabled",
      "budget_tokens": 10000
    },
    "messages": [
      {
        "role": "user",
        "content": "Are there an infinite number of prime numbers such that n mod 4 == 3?"
      }
    ]
  }'

Response Format

When Extended Thinking is enabled, the response includes two content-block types: thinking and text:

json

{
  "content": [
    {
      "type": "thinking",
      "thinking": "Let me analyze this step by step...",
      "signature": "WaUjzkypQ2mUEVM36O2TxuC06KN8xyfbJwyem2dw3URve..."
    },
    {
      "type": "text",
      "text": "Based on my analysis..."
    }
  ]
}

Thinking summary

Claude 4 models return Summarized Thinking rather than the full reasoning content. This helps preserve reasoning capability while preventing abuse. Claude Sonnet 3.7 still returns full thinking output.

Streaming Responses

When streaming, thinking content is returned via the thinking_delta event:

python

import anthropic

client = anthropic.Anthropic(
    api_key="<your ZENMUX_API_KEY>",
    base_url="https://zenmux.ai/api/anthropic"
)

with client.messages.stream(
    model="anthropic/claude-sonnet-4.5",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000
    },
    messages=[
        {"role": "user", "content": "What is 27 * 453?"}
    ]
) as stream:
    for event in stream:
        if event.type == "content_block_delta":
            if hasattr(event.delta, "thinking"):
                print("Thinking:", event.delta.thinking, end="")
            elif hasattr(event.delta, "text"):
                print("Answer:", event.delta.text, end="")

Google Vertex AI API (Thinking Mode)

Google Vertex AI Gemini models provide reasoning capabilities via Thinking Mode, controlled with the thinking_config parameter.

Supported Models

Gemini 3 Pro / Pro Image
Gemini 2.5 Pro / Flash / Flash-Lite

Parameter Reference

Gemini 3 and later use thinking_level, while Gemini 2.5 uses thinking_budget:

Gemini 3+ (thinking_level)

python

thinking_config=types.ThinkingConfig(
    thinking_level=types.ThinkingLevel.HIGH  # LOW / HIGH
)

Gemini 2.5 (thinking_budget)

python

thinking_config=types.ThinkingConfig(
    thinking_budget=10000  # number of tokens; -1 means dynamic budget
)

Model	Min Budget	Max Budget
Gemini 2.5 Flash	1	24,576
Gemini 2.5 Pro	128	32,768
Gemini 2.5 Flash-Lite	512	24,576

TIP

For Gemini 2.5 Flash/Flash-Lite, set thinking_budget=0 to disable thinking
For Gemini 2.5 Pro and Gemini 3 Pro, thinking cannot be disabled

API Call Examples

Python - Gemini 3Python - Gemini 2.5TypeScript

python

from google import genai
from google.genai import types

client = genai.Client(
    api_key="<your ZENMUX_API_KEY>", 
    vertexai=True,
    http_options=types.HttpOptions(
        api_version='v1',
        base_url='https://zenmux.ai/api/vertex-ai'
    ),
)

response = client.models.generate_content(
    model="google/gemini-3-pro", 
    contents="Find the race condition in this multi-threaded code snippet...",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_level=types.ThinkingLevel.HIGH
        )
    ),
)

print(response.text)

python

from google import genai
from google.genai import types

client = genai.Client(
    api_key="<your ZENMUX_API_KEY>", 
    vertexai=True,
    http_options=types.HttpOptions(
        api_version='v1',
        base_url='https://zenmux.ai/api/vertex-ai'
    ),
)

response = client.models.generate_content(
    model="google/gemini-2.5-pro", 
    contents="Solve the equation: x^3 - 6x^2 + 11x - 6 = 0",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_budget=16000
        )
    ),
)

print(response.text)

import { GoogleGenAI } from "@google/genai";

const client = new GoogleGenAI({
  apiKey: "<your ZENMUX_API_KEY>", 
  vertexai: true,
  httpOptions: {
    baseUrl: "https://zenmux.ai/api/vertex-ai", 
    apiVersion: "v1",
  },
});

async function main() {
  const response = await client.models.generateContent({
    model: "google/gemini-2.5-pro", 
    contents: "Solve the equation: x^3 - 6x^2 + 11x - 6 = 0",
    generationConfig: {
      thinkingConfig: {
        thinkingBudget: 16000, 
      },
    },
  });

  console.log(response.text);
}

main();

Viewing the Thinking Process

In Vertex AI Studio, you can view the model’s full thinking process. Via the API, thinking content is included in the response:

python

# Get the thinking summary (if available)
if hasattr(response, 'candidates') and response.candidates:
    candidate = response.candidates[0]
    if hasattr(candidate, 'thinking_metadata'):
        print("Thinking summary:", candidate.thinking_metadata.thought_summary)

Protocol Comparison

Feature	Chat Completion	Responses API	Anthropic Messages	Vertex AI
Parameter name	`reasoning_effort` / `reasoning`	`reasoning`	`thinking`	`thinking_config`
Intensity	`low`/`medium`/`high`	`effort` field	`budget_tokens`	`thinking_level`/`thinking_budget`
Token budget	`reasoning.max_tokens`	❌	`budget_tokens`	`thinking_budget`
Reasoning summary	`reasoning` attribute	`summary` field	`thinking` block	Thinking summary
Streaming	✅	✅	✅ `thinking_delta`	✅
Can be disabled	✅ `enabled: false`	✅ `effort: "none"`	❌ some models	❌ some models

Reasoning Models ​

OpenAI Chat Completion API ​

Parameter Reference ​

reasoning_effort ​

reasoning ​

effort ​

max_tokens ​

enabled ​

Parameter Priority and Auto-Calculation ​

Default parameter completion ​

Automatic max_tokens calculation ​

Reverse inference of effort ​

API Call Examples ​

Native call with the OpenAI Python SDK ​

Advanced usage with the OpenAI Python SDK ​

cURL ​

Using the reasoning_effort parameter ​

Extracting reasoning output ​

OpenAI Responses API ​

Parameter Reference ​

API Call Examples ​

Response Format ​

Anthropic Messages API (Extended Thinking) ​

Supported Models ​

Parameter Reference ​

API Call Examples ​

Response Format ​

Streaming Responses ​

Google Vertex AI API (Thinking Mode) ​

Supported Models ​

Parameter Reference ​

API Call Examples ​

Viewing the Thinking Process ​

Protocol Comparison ​

Reasoning Models

OpenAI Chat Completion API

Parameter Reference

reasoning_effort

reasoning

effort

max_tokens

enabled

Parameter Priority and Auto-Calculation

Default parameter completion

Automatic max_tokens calculation

Reverse inference of effort

API Call Examples

Native call with the OpenAI Python SDK

Advanced usage with the OpenAI Python SDK

cURL

Using the reasoning_effort parameter

Extracting reasoning output

OpenAI Responses API

Parameter Reference

API Call Examples

Response Format

Anthropic Messages API (Extended Thinking)

Supported Models

Parameter Reference

API Call Examples

Response Format

Streaming Responses

Google Vertex AI API (Thinking Mode)

Supported Models

Parameter Reference

API Call Examples

Viewing the Thinking Process

Protocol Comparison