Transform workbooks
Restructure sheets, add calculated columns and keep numbers as real numeric cells.
CLAUDE SANDBOX
The Claude sandbox is Anthropic’s code-execution environment, where code runs in Python and Bash on your files. Use it with Claude, or keep chatting with ChatGPT while Claude’s sandbox does the work behind the scenes.
Included with Pro, Smart, Team Pro and Team Smart plans. Uses your shared AI allowance.
Runs in chats with
“Compare our Q4 budget with actual spend and flag every line that is over by more than 10%.”
$ python compare_budget.py budget_q4.xlsxvariance = actual / budget - 1flagged = lines[variance > 0.10]wb.save("q4-variance.xlsx")4 of 22 budget lines are more than 10% over. Travel is the largest, at 27%.
WHAT YOU CAN DO
Claude’s sandbox is a strong fit for multi-step file work: transforming workbooks, checking calculations and building documents.
Restructure sheets, add calculated columns and keep numbers as real numeric cells.
Calculations are run and checked in code, so answers are based on observed results.
Build charts from your data, then see them in the answer and as downloadable images.
Create Word, PowerPoint and PDF files from data, notes or an existing file.
Scripts and command-line steps for jobs that go beyond a single formula.
Prefer ChatGPT’s style? Chat with ChatGPT and let Claude’s sandbox run the code.
Downloads can include XLSX · DOCX · PPTX · PDF · CSV · TSV · JSON · TXT · Markdown · PNG · JPEG · WebP · GIF. Office files are macro-free, and every file is security-checked before you can download it.
BEST OF BOTH
Pair ChatGPT’s conversation with Claude’s sandbox. ChatGPT stays your chat model: when a request needs code, it hands a precise task to Claude’s sandbox, gets the observed results back and explains them to you.
THREE STEPS, ONE CONVERSATION
Claude chats use this sandbox automatically. In a ChatGPT chat, choose Claude sandbox under Change file processing.
Upload a workbook or document and say what you want done.
The sandbox runs the code, and your chat model explains the result and offers the downloads.
MIX AND MATCH
Your chat model and your sandbox are separate choices. ChatGPT can even run its code in Claude’s sandbox.
| Chat | ChatGPT sandbox | Claude sandboxThis page | MultipleChat tools |
|---|---|---|---|
| ChatGPT chat | AvailableAutomatic default | AvailableClaude runs the code | Available |
| Claude chat | Not available | AvailableAutomatic default | Available |
| Gemini chat | Not available | Not available | AvailableAutomatic default |
| Grok chat | Not available | Not available | AvailableAutomatic default |
| Smart mode or several models | Not available | Not available | Available |
| Free plan | Not available | Not available | AvailableIncluded |
Provider sandboxes work in single-model chats for signed-in users on Pro, Smart, Team Pro and Team Smart plans. Everything else uses MultipleChat tools.
YOU’RE IN CONTROL
Leave it on Automatic and MultipleChat picks the native sandbox for your model. Or choose one yourself, for a single answer or as your default.
Open File preferencesApplies to that chat only. Rerunning uses more allowance.
Syncs across your devices.
THREE WAYS TO WORK WITH FILES
All three let your AI work with files. They differ in which models they support, who runs the code and what is included in your plan.
OpenAI’s Code Interpreter for ChatGPT chats.
Anthropic’s code execution for Claude and ChatGPT chats.
Built-in file tools for every model, on every plan.
KNOW WHAT TO EXPECT
Included with Pro, Smart, Team Pro and Team Smart plans. Sandbox work uses your shared AI allowance, depending on the model, task and tools.
Free plans use MultipleChat tools. Compare plans ↗
Eligible original files go to Anthropic, which runs the sandbox. With ChatGPT, OpenAI receives the results.
Access, size and security checks apply. ZIP uploads are not supported; upload the individual files instead.
Web search is off while a provider sandbox is selected. Smart mode and multi-model chats use MultipleChat tools.
Workspaces are temporary. If one resets, saved files are reused, but unsaved working state is lost.
Yes. Select Claude sandbox in a ChatGPT chat. ChatGPT remains the model you talk to, and Claude’s sandbox runs the code. Usage counts for both providers.
No. The ChatGPT sandbox runs in ChatGPT chats only. Claude chats can use the Claude sandbox or MultipleChat tools.
No. Describe what you want in your own words. The AI writes and runs code when that helps, then explains the result in plain language.
Below an answer, open Change file processing, choose another option and select Run again. That change applies to the conversation only. To change the default for future chats, open Account → File preferences → Default file processing. A choice you make in a conversation takes priority over your account default.
On eligible plans, ChatGPT chats use the ChatGPT sandbox, Claude chats use the Claude sandbox, and Gemini and Grok use MultipleChat tools. Free plans always use MultipleChat tools. If a sandbox chosen automatically cannot be confirmed, the chat continues with MultipleChat tools rather than stopping.
Provider sandboxes are included with Pro, Smart, Team Pro and Team Smart plans when you have AI allowance available. Sandbox work uses your shared allowance; how much depends on the model, the task and the tools used. Rerunning an answer with another sandbox uses more allowance. MultipleChat tools are included on every plan. Compare plans →
Not while the ChatGPT or Claude sandbox is selected. Choose MultipleChat tools for web research. Provider sandboxes also work in single-model chats only; Smart mode and multi-model chats use MultipleChat tools.
Eligible original files are sent to the provider that runs the sandbox: OpenAI for the ChatGPT sandbox and Anthropic for the Claude sandbox. File access, size and security checks apply, and every generated file is checked before you get a download. Sandbox workspaces are temporary: if one resets, saved files are reused when available, but variables and unsaved working files are lost.
MultipleChat connects you to OpenAI’s and Anthropic’s native code-execution tools. Results can still differ because of the model, instructions, available tools and execution limits. It is access to those capabilities, not a copy of every feature in the providers’ own apps. Review important calculations and files before relying on them.
Analyze data, run code and create downloads, with the AI you prefer.