NotebookLM Code Execution: your notebook now runs the code, not just reads it
As of June 8, 2026, every notebook gets its own secure cloud computer. NotebookLM writes and runs code — usually Python — inside an isolated sandbox to compute the numbers in your sources instead of estimating them. This is the change that turns a document reader into a data analyst: exact arithmetic, real data parsing, and charts generated from your own files. Here's what it does, when it fires, and who can use it right now.
Yes. Since June 8, 2026, every NotebookLM notebook includes a secure, sandboxed cloud computer that writes and executes code (typically Python) to analyze your sources — exact arithmetic, data parsing, and generated charts, computed rather than estimated. It’s live first for Google AI Ultra and Workspace tiers, on the web.
TL;DR — Each NotebookLM notebook now has a sandboxed cloud computer that writes and executes code (typically Python) to analyze your sources: precise arithmetic, dataset parsing, statistics, and generated charts — computed, not guessed. It runs on Gemini 3.5 + Antigravity with 100+ curated skills the system auto-selects. Live first for Google AI Ultra and Workspace business customers (AI Ultra / AI Expanded access), web-only, rolling out to other tiers over time.
Verified June 8, 2026. Maintained by a small team of power users focused on multi-AI research and learning workflows — no affiliate relationship with Google. See the full June update →
What actually shipped
A per-notebook secure cloud computer — the most technically significant part of the June 8 upgrade
On June 8, 2026, Google moved NotebookLM onto Gemini 3.5 and the Antigravity agentic framework, and gave every notebook its own secure, sandboxed cloud computer. That sandbox is what enables code execution: NotebookLM can now write code, run it in an isolated environment, read back the result, and use it in its answer.
It ships with 100+ curated software skills spanning analysis, calculations, file conversion, and other structured tasks. You don't pick which skills to switch on — NotebookLM decides which are relevant to what you're asking and orchestrates them itself. That's the "agentic" part: the model plans the steps, runs the tools, and (thanks to Gemini 3.5's visible thinking) shows you the intermediate work.
What it actually does
The difference between describing a number and computing it
Until June 8, NotebookLM was a retrieval tool: it read the pages you gave it and answered by citing them. If your source said revenue was $4.2M in 2024 and $5.1M in 2025, and you asked for the year-over-year change, the model produced a text estimate — and text estimates of math are exactly where language models slip.
Now it opens the sandbox, writes a short script, runs the calculation, and hands back a computed figure — and, if you want, a chart or a spreadsheet built from it. Concretely, the cloud computer can:
· Do exact arithmetic and statistics — percentages, growth rates, aggregates, ratios — grounded in the numbers in your sources.
· Parse and clean structured data — reconcile messy tables, normalize inconsistent rows, transform raw records into a usable shape.
· Generate charts and visuals — bar/line/trend graphs rendered from your data, not stock imagery.
· Build downloadable deliverables — the same run can emit a PDF report, an Excel (.xlsx) sheet, a PowerPoint (.pptx) deck, or .csv/.json.
The result is more trustworthy than a text-based guess because the number is calculated, and — with Gemini 3.5's visible reasoning — the steps are auditable.
Before vs after June 8, 2026
The same request, on the old RAG reader vs the new code-running notebook
| Dimension | Before — RAG reader | After — code execution |
|---|---|---|
| Math on your data | Text estimate, can hallucinate | Computed in a sandbox |
| Messy tables / datasets | Described, not processed | Parsed, cleaned, transformed |
| Charts | Not generated from data | Rendered from your numbers |
| Deliverables | Text + limited exports | PDF · .xlsx · .pptx · .csv/.json |
| How it works | Retrieve & summarize | Plan → write code → run → verify |
| Transparency | Loading spinner | Visible reasoning steps |
When it triggers — and how to invoke it
You don't flip a switch; you ask in plain language and NotebookLM decides to run code
There's no "code mode" button. The sandbox engages when your request implies computation over source data that's dense with numbers, tables, or logical structure. The reliable pattern is to state the calculation and the output you want in one prompt. For example:
"Analyze the three PDF financial statements in the sources, compute the year-over-year percentage change in revenue, and create a report with a bar chart."
NotebookLM will pull the figures from your sources, write and run a script to compute the deltas, render the chart, and assemble the report. The more explicit you are about the inputs ("the three PDFs"), the operation ("YoY % change"), and the artifact ("a report with a bar chart"), the more reliably it routes to the sandbox instead of answering in prose.
Where code execution earns its keep
The tasks that were painful or impossible on the old reader
Financial analysis
Compute growth rates and margins across statements, then export a chart-backed PDF or an .xlsx model.
Data cleaning
Turn inconsistent rows and messy tables from your sources into a structured, usable dataset.
Cross-source reconciliation
Line up numbers that live across several documents and check that they actually agree.
Trend & stats
Run aggregations and simple statistics, then render trend graphs from the computed values.
Unit & format conversion
Convert units, reshape data, and switch file formats using the curated skill library.
Report assembly
Package the computed results into a formatted, downloadable deliverable in one pass.
Which plans have it — and where
This is the catch: code execution is not on every tier yet
The June 8 agentic features — code execution, autonomous web sourcing, and the expanded exports — rolled out globally on the web first for two groups only: Google AI Ultra subscribers, and Workspace business customers with AI Ultra Access or AI Expanded Access. Free-tier users were not included at launch; Google said it will expand access over time but gave no schedule. The rollout is web-only — the mobile app wasn't named as a launch surface.
The underlying numeric caps (sources, daily chats, uploads) didn't change on June 8 — only the model and the feature set did. For the full per-plan breakdown of those caps and quotas, see the complete limits & benchmarks spec sheet →
Free prompt: make NotebookLM run the numbers
Copy this into an Ultra/Workspace notebook with numeric sources to force the sandbox path
The full Code-Execution prompt pack
12 battle-tested prompts that reliably route NotebookLM into the sandbox: financial models, dataset cleaning, cross-source reconciliation, statistical checks, and one-pass report/spreadsheet export — with the phrasings that stop it from answering in prose.
NotebookLM stopped guessing at your numbers and started computing them
- Computed beats estimated. Math done in code is verifiable; math done in prose is a guess. For finance, research, and anything with tables, that's the whole ballgame.
- One tool instead of five. Read, analyze, chart, and export in the same place — no exporting to a spreadsheet, running formulas, and pasting results back.
- Auditable by design. Gemini 3.5's visible reasoning shows the steps, so you can check the work instead of trusting a black box.
New capability, same discipline: verify the inputs, then let it run ↓
Frequently asked questions
Can NotebookLM run Python now?
Yes. As of June 8, 2026, each notebook has a secure, sandboxed cloud computer that writes and executes code — typically Python — to analyze the data in your sources. It runs the code in an isolated environment and uses the computed result in its answer, rather than estimating in text.
Is NotebookLM's math actually accurate now?
More reliable, yes. Because calculations run as executed code instead of language-model text prediction, arithmetic and statistics are computed rather than guessed. Gemini 3.5 also surfaces its reasoning steps so the work is auditable. You should still verify that the input figures were extracted from the right sources.
Which plans have code execution?
At launch on June 8, 2026 it's live for Google AI Ultra subscribers and Workspace business customers with AI Ultra Access or AI Expanded Access, on the web. Free-tier users were not included initially; Google said it will expand access over time but gave no schedule.
How do I make NotebookLM run code instead of just answering?
There's no manual toggle. State the calculation and the deliverable explicitly in one prompt — e.g. "compute the year-over-year revenue change from the three PDFs and create a report with a bar chart." Requests that name the inputs, the operation, and the output artifact reliably route to the sandbox.
What can it output?
The same run can produce computed answers plus downloadable deliverables: PDF reports with charts and tables, Excel (.xlsx) spreadsheets, PowerPoint (.pptx) decks, Markdown, and structured .csv/.json — part of the 11 new export formats in the June update.
Is my data private and sandboxed?
Each notebook's computer is isolated, and Google states that personal data uploaded to NotebookLM is not used to train its models. Workspace and Enterprise editions add stronger data protection. For confidential or privileged material, review your plan's terms before uploading.
Does it work on the mobile app or free tier?
Not at launch. The June 8 rollout is web-only and limited to Ultra and qualifying Workspace tiers. The mobile app was not named as a launch surface, and free/Plus/Pro access is expected to follow over time without a published date.