NotebookLM Is Now Gemini Notebook: What Actually Changed

First question everyone’s asking today: “Did my NotebookLM notebooks just disappear?” No. They didn’t. Your notebooks are still there, your URLs still work, and if you had a shared link floating around Slack, it still opens. The rename is real, the plumbing underneath is mostly the same, but there’s one genuine new feature that matters – and it’s not available to everyone yet.

Here’s the short version: Google announced on July 16, 2026 that NotebookLM is now Gemini Notebook. It’s still a standalone product. The Workspace Updates blog confirms automatic redirects keep existing links working. Mobile users may need to update the app. That’s the housekeeping.

The rename is a distraction – the real change is a cloud computer

Every article today is arguing about the name. Skip that. The substantive change is this: each notebook can now spin up a secure, isolated cloud environment that writes and executes Python against your sources. Google’s own framing calls it a “secure cloud computer,” and it’s what turns Gemini Notebook from a reading companion into something closer to a small data workbench.

The catch: on day one, this only works for Google AI Ultra subscribers and Workspace business customers with AI Ultra Access or AI Expanded Access. Pro users on the web get it over the coming weeks. Free users get the new name and logo and nothing else for now. So if you upgraded expecting instant code execution, check your tier first.

Community reactions: mostly “finally” with a side of confusion

The reception is split. Some people think dropping “LM” was overdue – Android Headlines called it Google killing an awkward name. One MakeUseOf writer in particular argued the old name told you exactly what the tool did and didn’t do: notebooks + a language model, boundaries obvious. Now those boundaries are getting blurrier.

Pro tip: If you already have Notebooks-in-Gemini and Gemini Notebook as separate mental models, keep them separate. Opening the same notebook inside the Gemini chat app lets it browse the web and pull from general knowledge. Opening it in the standalone Gemini Notebook keeps it strictly grounded to your sources. Same notebook, two behaviors. Pick the surface based on how much freedom you want the model to have.

How to actually use the new code execution (step by step)

Assuming you’re on a tier that has it, the workflow looks like this. There’s no “turn on code execution” toggle – it’s automatic when your prompt implies data work.

1. Open notebooklm.google.com (the URL still works – you’ll see the new blue/purple Gemini gradient logo).
2. Create a notebook and upload a data source. CSV works well for structured data. Also accepted as of mid-2026: PDFs, Google Docs, Sheets, images with text, and audio transcripts.
3. In chat, ask something that requires computation. Not “summarize this” – that doesn’t trigger code. Try: “Compute the median monthly spend by category and plot the top 5 as a bar chart.”
4. Watch the response. The model will write Python, execute it in the cloud sandbox, and return the chart plus a written explanation grounded in your file.
5. If the chart is wrong, ask it to show its code and re-run with a fix. It’ll iterate against the actual data instead of producing a plausible-looking answer from memory.

One hard limit carried over from Google’s broader code execution stack: per Vertex AI docs, code runs can hit a 30-second timeout. For notebook-scale analysis that’s fine. For anything heavier, you’re in wrong-tool territory – export the data and run it in Colab.

Which raises a question nobody has answered cleanly yet: what happens to the code execution sandbox if you have 300 sources loaded and the retriever pulls context from the wrong 10? The Python runs fine – it just runs on bad inputs. That gap between “code executed” and “code executed correctly” is going to be the source of a lot of confused support tickets over the next few months.

A real example: turning a research PDF stack into a comparison table

Say you’ve dumped ten market-research PDFs into a notebook. Before this update, you’d ask questions and get citations back. Now you can ask: “Extract every revenue figure mentioned in these ten sources, tag it with the source and year, and build a table sorted by year.”

The model will build a dataframe, run the extraction, and return a table you can copy. Whether every row is correct still depends on what’s in your sources – code execution doesn’t fix RAG retrieval quality. Speaking of which:

The gotcha that hasn’t gone away

Per-source caps didn’t budge. Every tier is still limited to 500,000 words or 200MB per source, whichever hits first (confirmed across multiple sources as of mid-2026, including Google help references). Ultra doesn’t lift that ceiling. If you have a 700,000-word thesis PDF, you still have to split it.

Answer quality also degrades as you approach the source cap on any tier. XDA’s writeup attributes this to how RAG retrieval works – the more chunks the retriever has to search, the noisier the “top matches” become. Adding code execution on top doesn’t fix retrieval. If the model can’t find the right passage, it will write Python against the wrong passage. Confidently.

There is something almost funny about that last point. Google added a cloud computer to a tool whose core limitation was always about finding the right text in the first place. Code execution makes the output more structured. It doesn’t make the inputs more accurate.

Tier reality check

Prices and source caps compiled from community pricing trackers as of mid-2026 – Google restructures these periodically, so double-check inside your account before upgrading.

What to do this week

Ultra subscribers: open your most data-heavy notebook and try a computation prompt today. That’s the only tier where the feature is live. Pro users get it in weeks – keep an eye on the rollout, no action needed until then. Free users? Nothing changes operationally; only the badge is different. And if you’ve been sitting on the fence about Pro, code execution is a more concrete reason to move than anything the pre-rename Pro tier offered.

FAQ

Do my old NotebookLM links still work?

Yes. Automatic redirects. No action required on your end.

Why does the same notebook behave differently in the Gemini app versus Gemini Notebook?

They share the notebook but not the engine. Inside the Gemini app, your notebook is available as context but the model still has full web access and general knowledge on tap – it’ll happily go beyond your sources if it thinks that helps. Inside the standalone Gemini Notebook, the model stays boxed in and only answers from what you uploaded. One surface is for open-ended research, the other is for source-grounded work where you need answers to stay inside your documents. Pick accordingly.

Is Gemini Notebook worth paying for now?

Depends on your work. Spreadsheets, PDFs with tables, financial docs – Pro at $19.99/mo gets genuinely useful once code execution lands. Pure reading and Q&A? The free tier’s 50 sources per notebook still covers most single-project workflows.