Code Interpreter

核心 · Key Idea

In one line: A Code Interpreter is a sandboxed environment (Python by default) that lets the model write, execute, and inspect code. When math, files, charts, or data are involved, the model emits code, runs it, sees the output, then keeps reasoning — handing the unreliable "mental arithmetic" off to a reliable computer.

What it is#

The flow:

User: Which 5 cities have the highest sales in this CSV?

Model: [generates Python]
import pandas as pd
df = pd.read_csv("/mnt/data/sales.csv")
print(df.groupby("city")["amount"].sum().nlargest(5))

Execute: [runs in sandbox]
city
Beijing     1234567
Shanghai     987654
...

Model: Top 5 cities by sales are … (grounded in real output)

The model decides whether and what to write, the sandbox executes safely, and results feed back to the model.

Analogy#

打个比方 · Analogy

An LLM doing mental math = a gifted bullshitter running your accounts.
With Code Interpreter, you hand them a calculator and scratch paper — they actually compute when numbers show up, and the answer becomes reliable.

Key concepts#

SandboxSandbox

Isolated container / VM / WebAssembly. Model code runs there and cannot touch the host.

Stateful KernelPersistent kernel

Jupyter-style — variables persist across turns; no need to re-read files each time.

File I/OFile mount

User uploads are mounted into the sandbox (`/mnt/data/`); the model can read and write.

OutputOutput capture

stdout / exceptions / charts are all captured and fed back to the model.

How it works#

The model treats execution output as an observation and continues reasoning — the same loop as ReAct.

Practical notes#

Hand off math / data / files first. Accounting, statistics, format conversion, plotting, regressions — all should go through Code Interpreter, ten times more precise than chat-style guessing.
Sandbox security. Always disallow network, SSH, sensitive directories. Minimum permissions — data analysis only is enough.
Kernel reuse vs fresh. Long sessions reuse the kernel for speed, but watch for variable pollution. Multi-tenant products usually issue per-user kernels.
Timeouts + resource caps. CPU / memory / wall time / output size all need limits — the model can generate infinite loops that pin the host.
Pipe errors back faithfully. When an exception occurs, send the full traceback back to the model — it will fix the code and retry. This is where Code Interpreter really shines.

Easy confusions#

Code Interpreter

The model **improvises code** to solve the problem.
Infinitely flexible, also risky — needs a sandbox.

Plain Function Calling

A **fixed set of tools** the model can pick from.
Predictable, but every new capability needs a new tool to be coded first.

Code Interpreter

**Write + execute**: model runs code in the sandbox and reads the result.

Code Generation

**Write only**: hands the code to the user; doesn't run it.
Cursor / Copilot default mode.

What it is#

Analogy#

Key concepts#

How it works#

Practical notes#

Easy confusions#

Further reading#