NotebookLM just became an agent, not a reader

Gemini 3.5 is the model now answering your questions; Antigravity is Google's agent-first, AI-powered IDE — the runtime that lets NotebookLM write and execute code behind the scenes. Together Google says they deliver more accurate, more reliable answers with better visibility into how the system reaches them.

For everyday use, the practical change is twofold. First, large-document and multi-source reasoning is more dependable, which matters most when a notebook holds dozens of long PDFs. Second, NotebookLM now exposes its thinking steps directly in chat, so you can see the chain of reasoning rather than trusting a black-box answer — useful for catching where a synthesis went sideways before you cite it.

Antigravity also ships a catalog of more than 100 curated software skills — the discrete capabilities (data parsing, charting, format conversion, computation) the agent draws on when a task needs code rather than prose.

Every notebook now includes a secure, isolated cloud computer that NotebookLM uses to write and run code for deeper research and more complex analysis. In plain terms: it can do real data work on your sources instead of only describing them.

Where this earns its keep:

• Messy data, unified. Point it at datasets with conflicting international formats — dates, decimal separators, currencies — and it reconciles them programmatically rather than guessing in prose.
• Math on your actual files. It performs accurate calculations directly against source data, so a financial model or statistical summary is computed, not approximated.
• Charts and reports. It can generate visualizations and assemble them into a finished document in one pass.

The mental model: treat a notebook less like a chatbot and more like a junior analyst who has a sandboxed laptop, your sources, and the ability to show their work.

Yes. You can now open a blank notebook with nothing but a question or a loose idea, and NotebookLM's chat will guide you through building a source repository — using Google Search to find relevant, high-quality material and add it for you. This removes the old precondition that you gather and upload everything before the tool was useful.

Two scenarios Google highlights are exactly the ones researchers hit constantly: finding primary sources in other languages to bring in perspectives outside your default search bubble, and surfacing related works by an author you've just discovered. The chat proposes sources; you decide what makes the cut.

A capability this large arrives with real friction — the stuff Google's announcement doesn't lead with, and the reason not to point this at high-stakes work unsupervised yet.

The rollout started June 8, 2026 on the web for Google AI Ultra subscribers and Workspace business customers with AI Ultra access. Google says the capabilities will reach other tiers over time, but has not published a timeline or pricing for free and Plus users, and mobile timing is unconfirmed.

If you don't see the new features yet, that's expected — confirm your plan qualifies, check you're on the web app, and watch our update log, which we revise as access widens. To compare what each plan and competitor offers, see the AI tool comparison and our system limits & benchmarks.

Two Google technologies landed in NotebookLM around the same window, and they do very different jobs — worth separating cleanly, because they’re easy to conflate.

Antigravity is the agentic backend — the brain. It’s Google’s agent-first runtime, and inside NotebookLM it does the thinking: it breaks a large research task into steps, writes and runs code in the per-notebook cloud computer, cleans and cross-checks your sources, browses the live web to fill gaps, and shows its working in the chat. Hand it something sprawling — “unify these messy financial tables and chart the trend” — and it plans the steps, writes the Python, and executes it behind the scenes rather than just describing what you’d do. For a big enough task it can spin up several agentic passes — one gathering fresh sources from the web, another cleaning your uploads, a third cross-referencing and de-duplicating — then synthesise the results.

Nano Banana 2 Lite is the visual renderer — the editor. It’s Google’s fast, cost-optimised image-and-video model, and its job lives in the Studio’s video features. Once Antigravity has done the reasoning and distilled the key script, Nano Banana 2 Lite turns that into the visual layer: the illustrations, charts, styled frames, and the roughly 60-second vertical clip of a Short Video Overview, animated and matched to AI narration.

The clean mental model: Antigravity does the research and runs the code; Nano Banana 2 Lite makes it watchable. They’re complementary halves of the same 2026 upgrade, not competing systems — which is why one notebook can both run a data analysis and hand you a vertical video summarising it.