Code execution tool: sandboxed Python for models, run in the EU

Add a tool with type: "code_interpreter" to a POST /v1/responses request. When the model decides that code would help, it writes Python, the API runs it in a sandbox, and the model reads the output, fixes errors if the code fails and runs it again until it has what it needs. Then it answers in text.

Each run appears in output as a code_interpreter_call item with the code the model wrote and the container_id it ran in. Showing that code to users, or logging it, makes the answer checkable: anyone can see how a number was calculated.

The container field sets where code runs. {"type": "auto"} creates a fresh sandbox for the request, or reuses one from earlier turns in the same conversation. Add file_ids to make files available inside it: upload a CSV, Excel file or JSON first with the files API using purpose user_data, then pass its id.

The sandbox has Python with common data libraries for tables, statistics and charts. Files the code creates, such as a cleaned CSV or a chart image, are referenced in the answer as container file citations so you can download them. Containers expire after a period without use, and their files are deleted with them.

Data analysis: attach an export from your shop, CRM or finance system and ask questions in plain language, such as growth by category or the customers with overdue invoices. Calculations: loan schedules, unit and currency conversions with rates you provide, or tax examples, where exact arithmetic matters. Transformations: clean a messy spreadsheet, merge two files or reshape data into the format another system expects.

Combine it with reasoning models for multi-step analysis, and with structured outputs if a program reads the final figures. For recurring reports, keep the prompt fixed and only change the attached file.

The sandbox is isolated from your systems and from other customers, and by default it has no access to the internet, so code cannot send data elsewhere or install packages from the network. If code fails, the error is returned to the model, which usually corrects it; persistent failures end with an explanation in the answer.

Runs have limits on time, memory and file size that depend on your plan, and we confirm them with you. Very large files are better summarised or filtered before upload. Check results that matter: the arithmetic is exact, but the model can still choose the wrong column or misread the question.