Google's NotebookLM Now Browses the Web, Runs Code and Builds PowerPoints
NotebookLM gets Gemini 3.5, a sandboxed cloud computer, and 11 export formats — turning a document Q&A tool into an agentic research workstation
Takeaways
- Category shift: NotebookLM upgraded from a document Q&A tool to an agentic research workstation with code execution and web browsing.
- New engine: Now powered by Gemini 3.5 and Antigravity, Google's agentic orchestration framework first unveiled at I/O 2026.
- Cloud computer: Each notebook gets a sandboxed cloud computer with 100+ curated software skills for running real code on real data.
- 11 export formats: PDF, DOCX, PPTX, XLSX, CSV, JSON, SVG charts, and AI-generated images are all now downloadable directly from the studio panel.
- Agentic source discovery: Users can start from loose ideas; NotebookLM will find and propose relevant web sources via Google Search with user permission.
- Access limited: Rolling out now to Google AI Ultra and Workspace AI Ultra/Expanded Access subscribers; free-tier timeline unconfirmed.
NotebookLM just crossed a line it has never crossed before. Google's research assistant, which built its reputation on grounding answers strictly in your uploaded documents, can now write and execute code, browse the web autonomously, and hand you a finished PowerPoint deck when it's done. This is not a feature refresh. It is a category change.
Three pillars, one overhaul
The platform is now running on Gemini 3.5 alongside a new framework called Antigravity , the same agentic infrastructure that Google unveiled at its developer conference in May 2026. According to the official announcement from NotebookLM's product director Trond Wuellner and software engineer Usama Bin Shafqat, the changes are substantial across three distinct areas: the chat interface, the output system, and the process for starting research projects.
Antigravity is the AI agent platform Google introduced with the Gemini 3 family , it is the engine that lets NotebookLM plan its actions and surface the steps behind them. Think of it as the orchestration layer that decides when to call a web search, when to spin up a code interpreter, and when to stitch results into a formatted file.
The cloud computer is the real story
Each notebook is now equipped with a secure cloud computer, enabling NotebookLM to write and run code useful for helping you perform deeper research and more complex analysis. The system includes more than 100 curated software skills, unlocking a wide range of new capabilities to help you more deeply understand the sources in your notebook.
Why does this matter so much? Because language models are notoriously bad at math. A traditional language model, when asked to sum a column in a financial statement, estimates the answer on a statistical basis and sometimes gets it wrong. A model that runs code opens the file, actually sums the column, and gives you the exact value , with the option to show you the script it used. That shift from statistical guessing to deterministic execution is what separates a toy from a working data tool.
An isolated virtual machine lets NotebookLM execute Python-level data work , calculations, dataset normalization, statistical analysis , in the background, instead of guessing numeric trends through token prediction.
Eleven ways to export your work
NotebookLM offers expanded output generation capabilities, allowing users to create and download editable deliverables in formats like PDFs, Excel sheets, PowerPoint presentations, CSV/JSON files and visual content for diverse professional applications. The full list of new formats includes:
- Data visualizations and charts (PNG, SVG)
- Documents: PDF reports, DOCX, Markdown, plain text
- Images via Nano Banana (PNG, JPG, GIF)
- Structured data: CSV and JSON
- Microsoft Excel (XLSX)
- Microsoft PowerPoint (PPTX)
Users can also edit outputs after generation. Previously, NotebookLM produced content only inside the chat window itself , text responses and audio overviews. The gap between AI-assisted research and a shareable deliverable is now much smaller.
Start from a blank page, not a pile of PDFs
The third pillar addresses the biggest friction point in the old workflow: you had to already know what you were looking for. With the latest update, you can start a chat about a project with the app, and it will help you build the knowledge base by suggesting different sources using its research skills and Google Search. The feature could help users get primary sources in other languages or find new material from related authors.
With your explicit permission, NotebookLM will autonomously crawl the live web, unearthing primary materials, cross-language perspectives, and highly targeted reference works to anchor your project. You stay in control , the system suggests and adds sources that you can accept or reject, and all sources remain clearly attributed.
The numbers Google is putting behind it
Google published internal benchmark data alongside the announcement. In side-by-side comparisons against the previous system, the upgraded NotebookLM achieved an average win rate of over 65% across five core evaluation dimensions , a 15 percentage point margin above parity. Two areas stood out:
- Large document analysis: 69.9% win rate
- Advanced web research and source discovery: 78.2% win rate
Worth noting: these are Google's own internal assessments, not independently verified benchmarks. The evaluation methodology , including what constitutes a "win" and how evaluators were selected , was not described in detail in the public announcement.
What you can actually do with it now
Google laid out three concrete workflow examples in the announcement:
- Data analysts can combine datasets from multiple countries with conflicting formatting, ask NotebookLM to conduct web research for additional context, run code for accurate analysis, and produce a PDF report with charts , all in one session.
- Program managers can feed in complex technical specifications and get back a simplified guide, a slide deck, and a step-by-step roadmap for their team.
- Small business owners can upload raw sales data alongside ad spend figures and have NotebookLM calculate campaign ROI and generate outputs ready to inform expansion decisions.
The common thread: raw inputs in, polished artifacts out, without leaving the notebook.
The RAG era is ending
NotebookLM was a RAG tool , Retrieval-Augmented Generation, the technique that pairs the model's generation with a targeted search inside the uploaded documents. You provided the sources, it answered by citing those texts and nothing else. That architecture was trustworthy but limited. You could interrogate your documents, but you could not act on them.
The June 8 update is the first that fundamentally changes what kind of tool NotebookLM is. Moving from a document-grounded Q&A system to an agent that writes and runs code, browses the web autonomously, and produces formatted files represents a different category of capability. The broader implication for the field: grounding AI in user-supplied documents was always a workaround for hallucination. Code execution is a more robust solution for anything numerical or structured, and the combination of both is where the real reliability gains live.
Who gets it, and what it costs
Google AI Ultra subscribers and eligible Workspace customers get first access, with broader availability planned later. Premium NotebookLM capabilities remain restricted to paid Google AI and Workspace subscribers. Free-tier users are not included in the initial rollout, and the new agentic features are web-only at launch , the mobile app is not yet a delivery surface for these capabilities.
If you are already on Google AI Ultra, the upgrade is live now at notebooklm.google. If you are on the free tier, Google has said broader access is coming but has not committed to a timeline.