ChatGPT gained a firewalled Python sandbox for uploaded files

The setup

OpenAI's own announcement introduced Code Interpreter on 23 March 2023 as one of two plugins it hosted itself. Access required a waitlist. The company describes giving its models a Python interpreter in a sandboxed, firewalled execution environment with ephemeral disk space; a session persists for a conversation's length, with an upper-bound timeout, and supports uploading and downloading files. On 6 July 2023, according to OpenAI's own release notes, the company began rolling code interpreter out in beta to all ChatGPT Plus subscribers over the following week, reachable through a toggle under beta features.

What the documents show

The March announcement is a vendor's own description, not an independent audit, but it is specific: the interpreter runs in a persistent, notebook-style session, and the company names three use cases from its own initial studies, solving mathematical problems, data analysis and visualization, and converting files between formats. The July note adds a plainer description: it lets ChatGPT run code, optionally with access to uploaded files, to analyze data, create charts, edit files and perform math. The current help center article, reflecting documentation as retrieved 16 September 2026, still describes a stateful environment displaying data tables interactively, and recommends reviewing generated code before relying on results, evidence the company still treats sandbox output as something a person must check.

The friction

None of the three documents claims the sandbox can reach the open internet; the March announcement instead specifies a firewalled environment, and that framing does not appear in either help center page, both of which simply omit outside network access altogether. That silence matters for anyone expecting the model to cross-reference an upload against a live source. The July note also frames the launch as a beta reachable only through a settings toggle, not default-on, and the current article notes file limits vary by upload type, model, plan and remaining allowance, a variable a user cannot check without hitting it.

What changed in the work

Before this feature, getting a chart or a statistical test out of a spreadsheet meant writing the code yourself or describing the problem to someone who could. The documented capability collapses that into uploading a file and asking in plain language, with the model running code inside the session OpenAI describes. What the sources do not support is trusting the output unchecked: the help article's own advice to review code, outputs and assumptions is worth taking literally, and it is an editorial read to say the tool changed who drafts an analysis, not who is responsible for it.

• Does the current file-type list and size limit match what a specific plan and workspace actually allow?
• Is there any indication, in documentation for the exact product used, that the execution environment can reach outside services?
• What in a generated chart or statistic would a reviewer need to re-derive by hand to trust it?

The gap between a vendor's plugin description and a help center article, three years apart, is a useful check on how a feature's own documentation ages: the sandbox framing has held, but the interface around it has not stayed still.