All capabilities Data and analysis

Let the model write and run real code mid-answer

Instead of describing an analysis, the model executes it in a sandbox, reads the actual output, and corrects itself before answering.

DocsOfficially documentedTestNot tested by benchr Quick · API Docs checked September 1, 2026

Claude APIGemini APIClaude Opus 5Claude Sonnet 5+4

Overview

Runs Python, and on Anthropic's tool also bash, inside a sandboxed container with a filesystem it can read and write.
Turns arithmetic and data questions from a guessing task into a computed one.
Renders charts inline on Gemini, where Matplotlib figures come back as images in the response.

01Attach the code execution tool and upload the input file rather than pasting the data into the prompt.
02Ask for the artefact and the code that produced it, so the result is auditable.
03Keep each run small. Gemini's environment stops at 30 seconds of runtime and allows up to five automatic regeneration attempts after an error.
04Expect text and CSV to behave best; both vendors describe those as the well-supported inputs.

Anthropic's code execution is not eligible for zero data retention, and it is unavailable on Amazon Bedrock and Google Cloud.
On Claude Haiku 4.5, programmatic tool calling and REPL state persistence are unavailable, so newer tool versions behave like the 2025 one.
Gemini's sandbox caps runtime at 30 seconds and cannot return media artefacts other than inline Matplotlib output.
Skills running on the Claude API sandbox have no network access and cannot install packages at runtime.

benchr has not run this capability. No test has been published.

benchr test runs →

Source-check history

September 1, 2026DocsOfficially documentedFirst logged from both vendors' current tool pages, including the Haiku 4.5 exception and Gemini's 30-second cap.