Code Interpreter
A tool that lets the model write and actually run code, then read what it printed.
Think of it like
Giving the model a scratch computer — it can type a program, hit run, and see the output, not just imagine it.
Example
Asked to analyze a spreadsheet, the model writes Python to load the file, computes the averages, and reports the numbers the code produced.
How it actually works
Instead of doing arithmetic or data wrangling in its head — where it quietly makes errors — the model emits code that runs in a sandboxed environment and returns real output. This is a huge reliability win for anything computational, and it doubles as a general tool: if there’s a library for it, the model can use it. The sandbox is essential, since running model-written code is exactly as risky as it sounds.
For product teams
It turns a model that fumbles math into one that computes exact answers.
For engineers
A tool that executes model-generated code in an isolated runtime and returns stdout, errors, and artifacts as observations.
Related
- Sandbox Execution — The isolation that makes running its code safe.
- Tool Use — The general act of running things for the model.
- Agentic Coding — A close cousin aimed at writing software, not just computing.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome