Skip to main content
QwenCloud home page
Docs
Tool calling

Code interpreter

Run Python in a sandbox

The code interpreter lets the model write and execute Python code in a secure sandbox during a conversation. When the model encounters a task that benefits from computation — such as math, data analysis, or chart generation — it automatically generates code, runs it, and incorporates the result into its response.

Getting started

The following examples show a complete code interpreter request with output.
  • Responses API
  • Chat Completions
  • DashScope
import os
from openai import OpenAI

client = OpenAI(
  api_key=os.getenv("DASHSCOPE_API_KEY"),
  base_url="https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
)

response = client.responses.create(
  model="qwen3-max-2026-01-23",
  input="What is 12 to the power of 3?",
  tools=[{"type": "code_interpreter"}],
  extra_body={"enable_thinking": True}
)

print(response.output_text)
print(response.usage)
Response example
12 to the power of 3 equals **1728**.

Calculation:
12^3 = 12 x 12 x 12 = 144 x 12 = 1728

How it works

Once enabled, the model processes each request in stages:
1

Thinking

The model analyzes the request and determines how to solve the problem.
2

Code execution

The model generates and executes Python code in a sandbox.
3

Result integration

The model processes the execution results and determines next steps.
4

Response

The model generates a natural-language response.
Steps 2 and 3 may repeat multiple times within a single request.

Supported models

Responses API
  • Qwen-Max: qwen3.8-max, Qwen3.7-Max series
  • Qwen-Plus: Qwen3.7-Plus series, Qwen3.6-Plus series, Qwen3.5-Plus series
  • DeepSeek: deepseek-v4-pro-0813, deepseek-v4-flash, deepseek-v4-flash-0731, deepseek-v4-pro
  • GLM: glm-5.2
  • Qwen3.8 open-source series
Chat Completions API / DashScope
  • Qwen-Max (thinking mode): Qwen3-Max series (requires thinking mode)
  • Qwen-Plus: Qwen3.5-Plus series

Other models

These models also support Code Interpreter but may not perform as well.
  • Qwen-Flash: Qwen3.7-Flash series, Qwen3.6-Flash series, Qwen3.5-Flash series
  • Qwen3.8 open-source series (qwen3.8-2.4t-a95b)
  • Qwen3.6 open-source series (excluding qwen3.6-27b)
  • Qwen3.5 open-source series
  • DeepSeek: deepseek-v4-pro-0813, deepseek-v4-flash, deepseek-v4-flash-0731, deepseek-v4-pro
  • GLM: glm-5.2

Retrieve executed code

The DashScope API returns the code that the interpreter executed in the tool_info field of each streaming chunk. Each entry contains the code string and a type of code_interpreter: 
"tool_info": [
  {
    "code_interpreter": {"code": "123**21"},
    "type": "code_interpreter"
  }
]
When the model calls the interpreter multiple times in a single request, tool_info accumulates all calls. The total call count is available in usage.plugins: 
"plugins": {"code_interpreter": {"count": 2}}
The Chat Completions API does not return the executed code. Use the Responses API or DashScope API if you need to inspect the generated code.

Limitations

  • Streaming required (Chat Completions and DashScope): The code interpreter only works with streaming (stream: true). Non-streaming requests return an error. The Responses API does not have this limitation.
  • The code interpreter and function calling are mutually exclusive and cannot be enabled at the same time.
    Enabling both results in an error.
  • When the code interpreter is enabled, a single request triggers multiple model inferences. The usage field aggregates token consumption across all inferences.

Billing

The code interpreter is temporarily free, but it increases token consumption.
Code interpreter - QwenCloud