Technology changes extremely quickly, and AI is certainly no exception. NotebookLM, however, has always felt like a bit of an outlier here (and I say this in the best possible way). It's a tool that got upgrades here and there, sure, but it remained largely unchanged at its core for a long while. We were anticipating significant upgrades at I/O 2026, but the company decided not to bring it up by name at all (well, it did get a slight mention at the very end, tied to some Gemini/science announcement).

That silence certainly had some of us fearing the worst — that it was an early indication Google was getting ready to quietly shelve the tool, the way it tends to do. However, a few days ago, NotebookLM got what Google itself is describing as its biggest launch yet. After having used NotebookLM since its early Google Labs days and going hands-on with this latest update, I have some thoughts.

NotebookLM now shows you what it's thinking before it answers

It finally shows its work

Regardless of what your go-to AI tool is, you've likely noticed that it sometimes takes a bit longer to respond to you. At these times, the tool is usually "thinking" and most of them now let you tap to expand that reasoning and see the steps it's working through. Given that NotebookLM's typically advertised as a thinking partner, you'd expect it would be one of the first tools to offer this, but that hasn't been the case. Well, until now.

As announced via a post on Google's The Keyword blog, a huge part of the "massive" update is that you can finally follow NotebookLM's reasoning directly in chat, watching it work through a task step by step instead of just handing you a finished answer. To put this to a test, I pulled up a Notebook I'd made to cross-verify details for a review I was working on. While seeing NotebookLM reason through your request doesn't really directly change the final output you get, it does make the whole process feel far more transparent.

For instance, I watched it catch a mismatch in one of my sources and work to reconcile it, rather than quietly glossing over it in the final answer. Similarly, I also noticed it flag that some of my sources were actually about a closely related model (rather than the one I was reviewing). Instead of NotebookLM relaying the detail as fact, it worked to verify the model name and specs across my other sources before committing to anything.

Getting to see its thinking trail rather than the tool landing on a verdict and expecting me to take its word for it makes a real difference, especially when you're relying on it for something where accuracy actually matters. And for the use cases NotebookLM is typically recommended for, which include research and synthesis, I think this kind of visibility is exactly the sort of thing the tool should have had all along.

NotebookLM can now write and run code

Code, of all things

When I think of NotebookLM, the last thing I imagine is anything to do with code. Yes, I've used the tool for pairing-adjacent use cases and have paired it with tools like Claude Code and Antigravity, and I've even used it to learn programming. But NotebookLM itself writing and running code isn't something I ever expected it to do, and yet here we are. The biggest structural change in this update is that NotebookLM now runs on Gemini 3.5 and... Antigravity. More importantly, every notebook you create now comes with a "secure cloud computer." This essentially means there is an isolated, sandboxed environment tied to your notebook where NotebookLM can write code and actually run it.

Given that NotebookLM's been designed to work with your own sources from day one, the best it could typically do before, even if you fed it actual code files, was read through them and explain what they did — walk you through the logic, summarize the functions, that kind of thing. It couldn't actually run any of it. Now, with the cloud computer, it can. Usama, an engineer on the NotebookLM team, confirmed that this is a Linux VM, which gives the model a broad toolset to work with — it can analyze data, run scripts, graph trends, build spreadsheets, and even write LaTeX, all within that environment. So, for instance, if you give it a spreadsheet and ask it to run calculations and turn all of it into charts or a downloadable report, it can write code to actually carry that out rather than just approximating an answer from what it can read.

On Google's blog post about this announcement, the company states that the system ships with over 100 curated software skills, which are simply pre-built capabilities it can reach for mid-task. Per the team, this is just the starting set, and the skills, coding libraries, and artifact types are expected to grow as the team learns how people are using the new NotebookLM. In his X post, Usama explains that AI models have been improving at coding faster than at almost anything else, and that "using a computer" actually maps onto a huge chunk of everyday knowledge work.

Ultimately, giving NotebookLM the ability to use one dramatically widens what it can do for you. The catch, in his framing, is that this only works well if the tool has strong context to draw on, which he argues NotebookLM has been built around from day one. Pairing that long-standing context with computer use is how he describes the leap from a research tool to an actual collaborative partner.

This update also introduced a bunch of new output formats

The export menu grew up

Right in the introduction of this article, I mentioned that NotebookLM has gotten a bunch of upgrades here and there. And if you look back at most of them, a lot were new Studio outputs. We got Audio Overviews, then Video Overviews, then Cinematic Video Overviews, more report formats, Slide Decks, Infographics, and so on. With this update, you can ask NotebookLm to create outputs in "even more formats" right within the Chat interface.

You can ask it to create you data visualizations and charts in png/svg, Documents (PDFs, docs, markdown, text files), Images with Nano Banana, CSV and jSON files, Excel sheets, and PowerPoints. What's interesting is that these aren't the old one-click templated buttons sitting in the Studio panel. Instead, you ask for them in chat, and NotebookLM plans the output, writes the code to build it, and runs it. The finished file then lands in the Studio panel for you to download, and the output is completely editable! For instance, I asked it to generate a PDF report and a PowerPoint deck from a notebook I'd put together. Rather than just spitting out files, it laid out a full plan first (a multi-page technical report with a custom color palette and section breakdowns, and a six-slide editable deck) and asked for my approval before actually building anything.

There's also a neat cross-language touch here: you can give NotebookLM your directions in one language and have it generate the actual output in another, which is handy if you're, say, researching in one language but need the final deliverable in a different one. From my testing so far, I believe the chat-based outputs are the stronger of the two. They feel a lot closer to what you'd get if you asked something like Claude or Codex to write a script and build a report for you, whereas the Studio panel outputs still feel like templated quick-generates by comparison.

The part where NotebookLM stops feeling like NotebookLM

AKA GeminiBookLM

One thing I've always, always praised NotebookLM for is the fact that it's grounded entirely in your sources. Unlike a general-purpose chatbot that pulls from the open web (and occasionally makes things up in the process), NotebookLM only ever worked with what you actually gave it. Whatever it told you could be traced straight back to a source you'd uploaded, citation and all. That closed, you-control-the-inputs setup is exactly what made it feel trustworthy in a way most AI tools just don't. It's the whole reason it never felt like just another chatbot to me.

That's why this next update is the one I have the most complicated feelings about. It's what Google's calling more effortless research. Previously, NotebookLM was at its most useful when you showed up with your own sources already gathered and a decent handle on your project. Now, you can start with nothing more than a loose idea or a question, and NotebookLM will help you build out your source repository right there in the chat. Say you want primary sources in another language, or you're chasing down related works by an author you just stumbled on — it can lean on Google Search to find relevant, high-quality sources from the web and pull them straight into your notebook.

Now, credit where credit is due, this is still grounded at its core. You're still the one who decides what actually makes it into a notebook, Google's been clear that nothing gets added without your say-so, and everything stays attributed. However, it also feels like the tool is turning more into a research agent that's a little too eager to go off and search the web for you, even when you didn't ask it to. That's where my reservations kick in. The whole reason NotebookLM stood apart was that it wasn't trying to be Gemini. It stayed in its lane, working only with what you handed it. The moment it starts reaching for the open web on its own, it starts feeling a lot more like every other AI tool out there, and a lot less like the NotebookLM I fell for.

A slight side note, but something I'm absolutely not a fan of in the slightest with this update is the constant overuse of emojis, especially at the end of responses. I'm also not a fan of NotebookLM ending each response with a "would you like me to do XYZ for you."

This is something a lot of other AI chatbots do, and exactly the kind of habit I never wanted creeping into NotebookLM. It's a small thing on its own, but it feeds right back into my bigger worry: every one of these little behavioral tics nudges it further away from the focused, no-nonsense research tool it used to be and closer to just another chatty assistant.

In some ways, I think NotebookLM's biggest update traded what made it special for something it doesn't need. I think its ability to write and run code is cool, but also, does NotebookLM really need to be a code-running, web-scouring, chart-generating research agent? Some of these additions are genuinely useful, and I'll absolutely use them. But a part of me can't shake the feeling that the tool is busy expanding outward to do more and more, when its real strength was always in how focused and restrained it was.