Google Turns NotebookLM Into a Full Research Agent With Gemini 2.5
NotebookLM gets Gemini 3.5, a cloud code sandbox, agentic web search, and a full suite of exportable output formats — all rolling out to Google AI Ultra subscribers now
Takeaways
- Agentic upgrade: NotebookLM gains multi-step reasoning and can now plan, search, and execute research tasks autonomously.
- Cloud code sandbox: Each notebook gets a secure cloud computer with 100+ software skills to write and run code for analysis.
- New model backbone: Powered by Gemini 3.5 and Antigravity, with a 65%+ win rate over the prior system in internal evals.
- Rich output formats: Export to PDF, DOCX, PPTX, XLSX, CSV, JSON, SVG, PNG, and more directly from the Studio panel.
- Cold-start solved: Start from a vague idea; NotebookLM now uses Google Search to build your source repository for you.
- Availability: Rolling out now to Google AI Ultra subscribers ($99.99/mo+) and Workspace AI Ultra/Expanded Access customers.
NotebookLM just crossed a meaningful line. Google's research tool has always been praised for keeping answers grounded in your own documents, but that same constraint made it a passive tool: you had to bring the sources, do the searching, and then manually export anything useful. That changes today. Google is rolling out a sweeping upgrade that gives NotebookLM agentic capabilities, a live cloud compute environment, and a full suite of downloadable output formats.
From RAG tool to research agent
The upgrade delivers new agentic capabilities in chat and more advanced reasoning to tackle the most complex research projects. The core shift is architectural. This is a massive leap for NotebookLM , moving from a closed RAG system strictly locked to source files to giving the system native agentic capabilities and external tool execution changes everything. RAG (Retrieval-Augmented Generation) is a technique where the model only answers based on documents you explicitly provide, rather than drawing from its general training data. It keeps hallucinations low but limits what the system can do on its own.
What separates agentic AI from standard chatbots comes down to autonomy and multi-step reasoning. A regular AI assistant might summarize a document when asked. An agentic system identifies what information is missing, searches for it, synthesizes findings, flags contradictions, and proposes how to resolve them , all before you ask. That's the capability Google is now baking into NotebookLM.
What's actually new under the hood
NotebookLM is now running on Gemini 3.5 and Antigravity, providing even more accurate and reliable information along with better visibility into the thinking process. Antigravity is Google's agent-first integrated development environment, and its inclusion here is what enables the new code execution layer.
- Secure cloud sandbox: Each notebook is now equipped with a secure cloud computer, enabling NotebookLM to write and run code for deeper research and more complex analysis. The system includes more than 100 curated software skills, unlocking a wide range of new capabilities.
- Transparent reasoning: The updated chat experience shows expanded thinking steps directly in the chat, making the tool more transparent and giving you insight into exactly how it arrives at its answers.
- Web-assisted source building: You can start a chat about a project and NotebookLM will help you build the knowledge base by suggesting different sources using its research skills and Google Search. The feature can help users get primary sources in other languages or find new material from related authors.
- Rich output formats: NotebookLM can now output to data visualizations, charts, documents (PDF, Word, Markdown, text), images, structured data (CSV, JSON), Microsoft Excel, or Microsoft PowerPoint.
In side-by-side evaluations against the prior system, the upgraded NotebookLM achieved an average win rate of over 65% , a 15 percentage point margin above parity , across the top five core evaluation dimensions. The gains are sharpest where it matters most: the system demonstrated substantial improvements in large document analysis, securing a 69.9% win rate, and achieved exceptional performance in advanced web research and source discovery, reaching a 78.2% win rate against the prior baseline.
The blank-slate problem is solved
One of the biggest friction points with NotebookLM has always been the cold-start problem. Until now, the tool was mostly relying on files and sources uploaded by users. That method reduced AI hallucinations because the responses remained anchored in trusted material, but it also meant that users had to do the hard work of collecting references.
You can now start with loose ideas and questions. NotebookLM can guide you through building your source repository directly in your chat, making it even easier to get started on any project. Crucially, NotebookLM can now search the web and recommend sources automatically, reducing the manual work of gathering references. You still approve what gets added, so the grounding guarantee stays intact.
What this is good for , and where it still has limits
The official announcement outlines three concrete workflows that show where the new capabilities shine:
- Data analysts can combine datasets from multiple countries with conflicting formatting, ask NotebookLM to run code for accurate analysis, and get back a chart-filled PDF report.
- Technical program managers can feed in complex integration specs and get a simplified guide, slide deck, and step-by-step roadmap for the team.
- Small business owners can upload raw sales data alongside ad spend reports and have NotebookLM calculate campaign ROI and model expansion scenarios.
The tool still shines brightest when you already know which documents contain the answers and just need help synthesizing across them. NotebookLM is the perfect tool when you already know which documents contain the answers, those sources are in different formats or mediums, and you need the AI to stick to what's actually in the documents because the stakes are too high for hallucinations.
That said, some rough edges remain. While the idea of granting more autonomy to AI and cloud-based tools seems appealing, the crucial question is whether NotebookLM will consistently provide accurate answers without necessitating constant verification. Power users have also noted that if you have domain expertise on a topic, you can probably filter out low-quality sources better than NotebookLM, and the Deep Research tools in Gemini, ChatGPT, and Claude still perform better for pure open-web research tasks. And some users are pushing back on the Ultra-tier paywall, with one commenter noting that NotebookLM is the most useful product for learners, and restricting its advanced features to the Ultra tier does not make sense.
The competitive context
The timing is deliberate. As enterprise AI tools explode across the market, Google has been racing to differentiate NotebookLM from the pack of generic chatbots flooding knowledge workers' desktops. Microsoft has been aggressively pushing Copilot across its Office suite, while Anthropic targets knowledge workers with Claude's expanded context window. Google's response is to go deep on the research workflow specifically , a space where source fidelity and citation quality matter more than raw generation speed.
Google's advantage lies in integration , NotebookLM sits within the broader Google Workspace ecosystem, meaning these capabilities could eventually flow into Docs, Sheets, and Gmail. That's a moat no standalone research tool can easily replicate.
Access and pricing
The NotebookLM updates are rolling out globally on the web starting today to individuals with a Google AI Ultra subscription and Workspace business customers with AI Ultra Access and AI Expanded Access, and will expand to other users soon. NotebookLM has four pricing tiers: Free, Plus ($7.99/mo via Google AI Plus), Pro ($19.99/mo via Google AI Pro), and Ultra at either $99.99/mo (20TB Plan) or $200/mo (30TB Plan) via Google AI Ultra. The new agentic features are currently locked to the Ultra tier, with broader rollout timing unspecified.
For anyone doing serious multi-source research , synthesizing papers, analyzing datasets, building deliverables from raw material , this update closes a lot of the gap between what NotebookLM promised and what it could actually deliver. The cloud sandbox and agentic web search are the two features to watch most closely as they roll out wider.