Source: Code execution (Gemini API docs)
Gemini's built-in Python code execution tool: 30-second runtime, up to 5 self-retries, fixed libraries, matplotlib-only charts, no extra charge beyond tokens.
Contents
- What: official API documentation
- Publisher: Google
- Date: living document, read 2026-11-14
- Original: Code execution
Summary
Enabling {"type": "code_execution"} lets Gemini write and run Python in a Google-hosted environment and iterate on the results before answering. It is the lightest of the provider tools: short runs, no package installs, outputs returned inline.
Key claims
- Python only; maximum runtime 30 seconds per execution.
- On an execution error the model may regenerate code, up to 5 times.
- Fixed library set (numpy, pandas, matplotlib, scipy, scikit-learn, sympy, opencv-python, pillow, tensorflow and others); you cannot install your own.
- File input supported (from Gemini 2.0 Flash), inline or via the Files API; files larger than the context window error. Text and CSV work best.
- Only matplotlib renders graphs; outputs come back as inline data.
- No extra charge for the tool: intermediate tokens (prompt, generated code, execution results) are billed as input; final code, results, summary and thinking as output.
- Combinable with Grounding with Google Search, and on Gemini 3/4 models with custom function calling.
- Can degrade other output quality (e.g. creative writing); success varies by model. Network access is not documented.