Verdict: Google's June 2026 overhaul turned NotebookLM — now renamed Gemini Notebook — from a source-summarising chatbot into a small research workstation. Each notebook now runs on the Gemini 3.5 model, gets its own sandboxed cloud computer that writes and runs code, and exports eleven real file formats including editable PowerPoint, Excel, and PDF. For anyone whose job is turning a pile of sources into a report, deck, or spreadsheet, it cuts the "research → draft → reformat in another tool" chain down to one prompt. It is genuinely useful, and it is the easiest agentic AI surface to try without writing or running any code yourself — but it still needs a clear brief, source-grounding discipline, and a human review pass before anything it produces ships.
Last verified: 2026-08-02
- Product renamed from NotebookLM to "Gemini Notebook" on 16 July 2026 — same product, same notebooks, same URL.
- New agentic features ship first to Google AI Ultra and Workspace customers with AI Ultra Access or AI Expanded Access; rolling out to AI Pro ($19.99/mo) on the web over the following weeks. Free tier not yet confirmed.
- Eleven export formats: PDF, DOCX, Markdown, TXT, XLSX, PPTX, CSV, JSON, PNG, JPG, SVG.
- Pricing/limits are volatile — re-check before quoting in a deck or contract.
What actually changed in Gemini Notebook on 8 June 2026?
Three things shipped at once, and the combination is what matters — not any single feature.
- The model is now Gemini 3.5. Google upgraded NotebookLM from the prior Gemini 3 stack to Gemini 3.5 (also referred to in some Google communications as Gemini 3.5 Flash), the same model family it positions as its "most intelligent model for sustained agentic and coding performance." This is a measurable reasoning upgrade, not a marketing refresh — Google's own internal eval cited in third-party reporting puts the new system ahead of the old one in roughly 65% (about two out of three) of test tasks, with the biggest wins on long documents and multi-step research (Google, "Better research with NotebookLM," 8 June 2026;see also The Verge, 16 July 2026 and heise online, 9 June 2026 for independent corroboration).
- Each notebook now has its own secure cloud computer. A sandboxed environment is attached to every notebook that lets the model write and run code on your behalf — chart a dataset, normalise a CSV, compute a breakdown, build a file. This is the single biggest shift. Before, NotebookLM could only reorganise text from your uploaded sources; it predicted a plausible-looking answer. With the cloud computer, when the model hits a data request, it can hand the task to a real execution path and compute the result instead of guessing it. Google is deliberately vague about the isolation boundary (no published detail on container isolation, network policy, or audit logging), which matters if you handle sensitive data (heise online, 9 June 2026).
- Antigravity is the orchestration backbone. Gemini Notebook shares agentic infrastructure with Antigravity, Google's separate agent-first development platform. In plain terms: when you ask Gemini Notebook to do something that needs tool use or multi-step reasoning, it uses the same orchestration engine builders are coding against in Antigravity — but you never see, install, or touch the IDE (Google blog, 16 July 2026). If you want the deeper version of the same engine for building software, see our guide on setting up Antigravity 2.0 as an agent OS.
How much does Gemini Notebook cost, and which tier do you need?
Gemini Notebook is not sold on its own. Paid tiers are bundled into Google's broader AI subscription plans, and the June 2026 agentic features land on the higher tiers first. The free Standard plan keeps the original source-grounding features plus Audio and Video Overviews; the cloud computer, code execution, and the new export formats are gated to paid plans during the rollout window.
| Plan | Price (USD/mo) | Sources / notebook | Daily chats | Cloud computer + new exports |
|---|---|---|---|---|
| Standard (Free) | $0 (any Google account) | 50 | 50 | Not yet; original features only |
| Plus (via Google AI Plus) | $4.99 | 100 | 200 | Rolling out later |
| Pro (via Google AI Pro) | $19.99 | 300 | 500 | Rolling out over weeks |
| Ultra 20TB (via Google AI Ultra) | $99.99 | 500 | 2,500 | Yes — first to get it |
| Ultra 30TB (via Google AI Ultra) | $199.99 | 600 | 5,000 | Yes — first to get it |
| Workspace Business (AI Ultra/Expanded) | from $14/user | Plus-equivalent | Plus-equivalent | Yes — first to get it |
Sources: Google Gemini Notebook plans page and Google support: "Gemini Notebook upgrade". The AI Plus price was cut from $7.99 to $4.99 on 8 June 2026 to coincide with the same launch. The student offer (free 12-month AI Pro) closed in the US on 30 April 2026. Limits are vendor-stated and should be re-checked on the official plans page before quoting in a deck — see our cost-optimisation guide for how to think about tier trade-offs more broadly.
What can Gemini Notebook actually make now? (The eleven export formats)
This is the part that makes the workflow change real. Before the June update, the answer to "can NotebookLM produce a PowerPoint?" was a workaround — export a PDF and rebuild it by hand. Now, the Studio panel can hand back a real, editable file in eleven formats.
| Format bucket | What you get | Use it for |
|---|---|---|
| Documents | PDF, DOCX, Markdown, TXT | One-pagers, briefs, reports, hand-off text |
| Spreadsheets | XLSX | Numbers, breakdowns, editable budgets |
| Slides | PPTX | Decks you can open in PowerPoint or Google Slides |
| Structured data | CSV, JSON | Hand-off to engineering, data tools, or BI |
| Images | PNG, JPG, SVG | Charts, diagrams, cover graphics |
The files land in the Studio panel and you can download them or iterate on them with another prompt ("add a recommendations section," "chart the top 3 channels instead"). Google says more formats are coming. Source: Google blog, 8 June 2026; corroboration in heise online and Creative AI News.
The research-to-deliverable loop: the workflow nobody named
The video that prompted this article demonstrates a workflow it never names. We'll call it the research-to-deliverable loop — the smallest unit of work Gemini Notebook is now good at. It is five steps, and you can run it today on a free or paid notebook.
Step 1 — Build the source stack first
Gemini Notebook is still, fundamentally, source-grounded. Every answer is anchored to documents you upload. The cloud computer and the export formats amplify the quality of your sources; they do not replace them. Spend the first ten minutes feeding the notebook the actual material — PDFs, the Google Sheets you want it to chart, the press releases, the pricing pages. Garbage in, garbage out still applies; if a source is a scanned, low-resolution PDF, the output inherits that noise. (Source-grounding is what distinguishes Gemini Notebook from open-web agents; see when not to use AI for the decision rule on which tasks belong here vs. on an autonomous agent.)
Step 2 — Ask it to find what you're missing
This is the new agentic source-discovery step and it is the cheapest research upgrade in the tool. Once your notebook has anything in it, ask: "Based on what's in this notebook, what important sources am I missing?" The notebook reads what you already have, surfaces real gaps (market size, competitor pricing, a case study you never added), then goes to the web and brings back candidates. Nothing is added without your sign-off — you click import on what you want to keep. You can also deliberately ask it to find the opposing view: "find me strong arguments against doing this in-house" — useful when a draft is starting to feel like a one-sided echo chamber.
Step 3 — Ask it to compute, not summarise
This is where the cloud computer earns its keep, and it is the shift the video justifiably hypes. Pre-June, if you uploaded an ad-spend sheet and a signups sheet and asked NotebookLM "which channel is converting best," it produced a confident-sounding paragraph guessing at the answer. Now it can write and run code on your two files, compute the real conversion breakdown, and hand you a number that traces back to your rows. Then ask it to chart that breakdown — the file lands in the Studio panel as a PNG. Then ask it to reason over the chart in plain words, telling you which channel is working and which isn't. Each step is grounded in something it actually executed, not in a plausible-sounding prediction.
The same loop works on any numbers you track: weekly signups, monthly traffic, a year of keyword rankings. If you want to see whether publishing new content is moving your rankings, you can chart your keyword positions against your publication dates and read the relationship straight off the chart.
Step 4 — Ask for the deliverable in a real format
This is the step that closes the loop and the piece that was genuinely missing from NotebookLM before. Instead of copying text out of the chat and rebuilding it in PowerPoint or Excel, you ask for the file directly.
A few patterns that work well:
- "Build me a one-page PDF that explains [proposition] and ends with a clear call to action." — Real headings, real structure, a problem section, a benefit section, a CTA. You download it and you have a first draft, not a blank page.
- "Lay out the cost and growth numbers as an Excel file." — Real spreadsheet with editable cells, not a static screenshot.
- "Design a landing page outline for [proposition] that shows the value and turns visitors into leads." — A structure you can hand to a designer or developer.
- "Write this up as a short PDF report with the chart built in." — Finished report ready to send, built entirely from your own uploaded numbers.
The Studio panel is where outputs collect; you can iterate on any of them with a follow-up prompt ("add a section on risks," "reduce to 5 slides") before you download.
Step 5 — Human review the file before it ships
This step is in our loop, not in the source video's, and it is the one that takes a "stunning" demo into something you can ship. The cloud computer can produce files fast — that speed is the trap. Three specific checks before anything it produces leaves your hands:
- Trace every number. Open the PDF/deck/spreadsheet next to the source it came from. Gemini Notebook now computes instead of guessing — but it computes on what you uploaded, and the source can be wrong, stale, or itself a vendor claim. See our framework for fixing AI slop in generated content; the same scrutiny applies to AI-generated files, not just AI-generated text.
- Check the call to action and the verdict. Templates the model reaches for (problem / benefit / CTA) tend toward a generic, affirming shape. If the notebook doesn't have the evidence for a strong claim, the deck will imply one anyway — replace it with what your sources actually support.
- Strip anything the model knows from outside your sources. Gemini Notebook is supposed to ground in what you uploaded, but the cloud computer and web search broaden the leak surface. A confident stat in the deck that you can't tie back to a source in your notebook is a fabrication risk, even when the underlying fact is true.
What this means for you
If you are a small business owner, a one-person marketing team, or a consultant producing client-facing reports, the simplest change to make this week: pick one recurring research-to-deliverable handoff — the weekly client PDF, the monthly board one-pager, the ad-spend breakdown deck — and try running it end-to-end inside a single notebook. Compare the time it took against your current chain (which is probably: research in one tool, draft in a second, format in a third). The agentic features are gated to paid tiers during rollout, so if you're on the free Standard plan today, the rest of NotebookLM — Audio Overviews, source-grounded chat, study guides — already works for free, and the upgrade is mostly about removing the formatting bottleneck at the end of that loop.
If you run an SEO or content operation, the implication is more specific. Good content starts with good research; good research now produces finished artefacts. Turning a notes marinade into a published, ranked page still needs the discipline we cover in the social SEO multi-platform ranking guide — the tool speeds the research step, not the ranking step. The quick-and-dirty deck Gemini Notebook builds from your research is a first draft, not a pillar page; the ranking work still happens after the export.
If you want a wider view of everything Google shipped around the same rebrand — the avatars, the music tools, and the Notebook rename itself — see our roundup of the July 2026 Google Gemini AI updates.
FAQ
Q: Is Gemini Notebook the same thing as NotebookLM?
A: Yes. Google renamed NotebookLM to "Gemini Notebook" on 16 July 2026. It is the same standalone product at the same URL (notebooklm.google); existing notebooks, shared links, and bookmarks continue to work without any action. The rebrand integrates the product more tightly into the Gemini app ecosystem, but the underlying tool is unchanged (Google blog, 16 July 2026).
Q: Does Gemini Notebook cost money, and is the free tier still useful?
A: The free Standard tier ($0, any Google account) keeps the original source-grounding features: up to 100 notebooks, 50 sources each, 50 chats a day, plus Audio and Video Overviews. Paid tiers ($4.99 Plus, $19.99 Pro, $99.99+ Ultra) raise the daily caps and unlock the new agentic features — cloud computer, code execution, and the 11 export formats — during the rolling rollout. Pricing is bundled into Google AI subscriptions, not sold standalone (Google Gemini Notebook plans).
Q: Can Gemini Notebook actually build a PowerPoint file I can edit?
A: Yes. PPTX is one of the eleven supported export formats. The file lands in the Studio panel, downloads as a real .pptx, and opens in PowerPoint or Google Slides with editable text, shapes, and notes. It is still a first draft — design polish, brand formatting, animations, and complex visual storytelling still belong in PowerPoint or Slides — but it is no longer a static image or PDF workaround (heise online, 9 June 2026).
Q: What is the "secure cloud computer" in Gemini Notebook, and is my data safe in it?
A: It is a sandboxed execution environment attached to each notebook that lets the model write and run code — to chart a dataset, clean a CSV, compute a breakdown, or generate a file — on your behalf. Google has not published technical detail on container isolation, network boundaries, or audit logging, which means a formal risk analysis is still your job before feeding it sensitive client or regulated data (heise online, 9 June 2026).
Q: How is this different from NotebookLM's Audio Overview feature?
A: Audio Overview (the podcast-style synthesis NotebookLM launched in 2024) is unchanged and still works on every tier — it turns your sources into a two-host "deep dive" audio discussion. The June 2026 upgrade is a different category of capability: the model can now act on your sources (run code, build files, find missing sources on the web) instead of only reading them to you. Audio Overview is the listening surface; the cloud computer is the working surface.
Q: What is the difference between Gemini Notebook and Google Antigravity?
A: Antigravity is a separate, standalone Google product — an agent-first software development platform with its own desktop app, CLI, and SDK, launched at I/O 2026 for builders who write code. Gemini Notebook shares some underlying agentic infrastructure with Antigravity (the cloud computer uses the same orchestration backbone) but remains a no-code research tool for non-developers. If you want the IDE version for building software, see our guide to setting up Antigravity 2.0 as an agent OS.
Discussion
0 comments