Every source remains clearly labeled, so when you're done, you know where your conclusions came from.
In the ongoing human effort to make sense of fragmented knowledge, Google has meaningfully expanded what its NotebookLM research tool can do — rebuilding its reasoning engine around Gemini 3.5 and granting it the ability to write and execute code within a researcher's own workspace. The upgrade arrives not as a novelty but as a response to how inquiry actually unfolds: messily, incrementally, across scattered sources and half-formed questions. What Google is offering is less a smarter chatbot and more a research environment that attempts to stay with the thinker through the entire arc of discovery, from loose fragments to finished deliverables.
- Research tools have long forced people to juggle five applications at once — NotebookLM is now attempting to collapse that fragmentation into a single environment.
- The addition of live code execution is the sharpest edge of this upgrade, letting the AI write scripts, run them in a secure cloud, and return real analytical results rather than approximations.
- Users had persistently demanded to see how the AI reasons, not just what it concludes — the new version surfaces its thinking step by step, giving researchers a chance to catch errors before they compound.
- Output formats have expanded from plain text to charts, PDFs, spreadsheets, PowerPoint decks, and structured data files, meeting researchers where their actual workflows live.
- The rollout is live for Google AI Ultra and Workspace subscribers, with broader adoption hinging on whether the reasoning proves trustworthy and the sourcing transparent enough to earn sustained use.
NotebookLM, the research assistant that has become essential to students, analysts, and knowledge workers over the past three years, is receiving its most substantial upgrade yet. Google has rebuilt its chat engine around Gemini 3.5 and added the ability to write and execute code directly within research notebooks — a change that fundamentally expands what the tool can accomplish.
The upgrade is designed around how research actually begins: not with a tidy library, but with fragments. A PDF, a spreadsheet, a handful of links, half-formed questions. NotebookLM can now help users build from that disorder — surfacing better sources through Google Search, keeping every source clearly labeled so conclusions remain traceable.
The code execution layer is the most consequential addition. Each notebook connects to a secure cloud environment where NotebookLM can write and run scripts, drawing on more than a hundred curated software skills. A researcher can upload messy datasets, have the tool clean and analyze them, generate charts, and export a finished PDF report without leaving the interface. A small business owner can combine sales and spending data and receive analysis that informs real decisions.
Reasoning transparency — something users had repeatedly requested — is now built in. When a complex question is posed, the AI shows its work as it unfolds, allowing users to catch errors earlier and place more confidence in the output.
Output options have expanded well beyond text: data visualizations, PDFs, Word documents, Markdown, CSV and JSON files, Excel spreadsheets, and PowerPoint presentations are all now available. The rollout is underway for Google AI Ultra subscribers and qualifying Workspace customers. Whether the tool earns broader adoption will depend on how reliably the reasoning holds and how much users trust the sources it surfaces — questions the coming months will begin to answer.
NotebookLM, the research assistant that has quietly become essential to millions of students, analysts, and knowledge workers over the past three years, is about to do a lot more than it did yesterday. Google has rebuilt its chat engine around Gemini 3.5, the company's latest reasoning model, and added something that changes what the tool can actually accomplish: the ability to write and execute code directly within your research notebooks.
The upgrade arrives as a recognition that research in the real world is messy. You don't start with a perfectly organized library of sources. You start with fragments—a PDF someone sent you, a spreadsheet of numbers, a handful of web links, half-formed questions. NotebookLM now helps you build from that chaos. The tool can guide you toward better sources, pull in results from Google Search, and let you decide what stays and what goes. Every source remains clearly labeled, so when you're done, you know where your conclusions came from.
What makes this version genuinely different is the code execution layer. Each notebook now includes access to a secure cloud computer where NotebookLM can write scripts, run them, and show you the results. The system comes loaded with more than a hundred curated software skills—think of them as pre-built thinking steps that let the AI tackle more sophisticated analytical work. A researcher can now upload messy international datasets, have NotebookLM write the code to clean and analyze them, generate charts, and produce a PDF report, all without leaving the interface. A technical team can take dense specification documents and convert them into simplified guides or presentation decks. A small business owner can combine sales figures with spending data and get analysis that actually informs whether to expand.
The reasoning transparency that Google emphasized is worth noting. Users asked for it repeatedly: they wanted to see not just the answer, but how the AI arrived at it. The new version shows its work. When you ask NotebookLM a complex question, you can now watch the reasoning unfold, which means you can catch errors earlier and trust the output more.
Output flexibility has expanded dramatically. You're no longer limited to text. NotebookLM can now generate data visualizations as PNG or SVG files, export documents as PDF, DOCX, Markdown, or plain text, create structured data in CSV or JSON format, and produce Microsoft files like Excel spreadsheets and PowerPoint presentations. Images can be generated through Nano Banana in standard formats. Google says more output types are coming.
The rollout is already underway for Google AI Ultra subscribers and Workspace business customers who have AI Ultra access. The timing matters: as AI tools proliferate and become more capable, the ones that actually integrate into how people work—rather than asking people to work around them—will be the ones that stick. NotebookLM has always been good at synthesis and summarization. Now it's becoming something closer to a full research environment, one that can handle the kind of work that used to require switching between five different applications. Whether that translates to the kind of adoption Google is betting on depends on whether the reasoning is actually transparent enough, whether the code execution is reliable, and whether people trust the sources it surfaces. Those are the questions the next few months will answer.
Notable Quotes
Users asked repeatedly for reasoning transparency—they wanted to see not just the answer, but how the AI arrived at it.— Google's product positioning
The Hearth Conversation Another angle on the story
What actually changes for someone who's been using NotebookLM for the past year?
The biggest shift is that you can now ask it to do analytical work that requires computation. Before, it was excellent at reading and summarizing. Now it can write code, run it, and show you the results—all inside your notebook.
So if I have a spreadsheet of messy data, it can clean it?
Yes. You upload the file, describe what you need, and it writes the script to do it. Then it runs that script and shows you the output. You're not exporting to Python or R anymore.
What about the reasoning transparency piece? Why did users care about that so much?
Because they were getting answers they didn't fully trust. If you can't see how the AI arrived at a conclusion, you have to either accept it or redo the work yourself. Showing the reasoning steps lets you catch errors and understand where the confidence is actually justified.
Is the source attribution actually reliable, or is that just marketing language?
It's built into the system—every source stays labeled throughout your research. But the quality of what it surfaces still depends on what you feed it and what Google Search returns. It's more trustworthy than a black box, but not infallible.
Who actually benefits most from this upgrade?
Anyone doing research at scale—academics, analysts, business teams. But also smaller operations that don't have dedicated data teams. A three-person startup can now do analysis that used to require hiring someone.
What's the catch?
You need AI Ultra access, which isn't cheap. And the code execution is only as good as the prompts you give it. It's still a tool that requires knowing what you're asking for.