Google's NotebookLM, one of the company's earliest generative AI experiments, is receiving its most significant update to date. The note‑taking and research assistant is moving to the latest Gemini 3.5 Flash model, gaining a new code‑execution feature called Antigravity, and expanding its output options to include multiple file formats. These changes aim to make NotebookLM faster, more accurate, and more useful for complex research tasks.
NotebookLM originally launched in mid‑2023 as a focused tool that allowed users to upload documents and web pages and then query them using Google's AI. It quickly stood out among the flood of AI chatbots because it was grounded in the user's own sources, reducing hallucinations and increasing reliability. The new update builds on that foundation by integrating the Gemini 3.5 Flash model, which debuted at Google I/O earlier this year. The Flash model is designed to deliver high‑quality responses with lower latency and reduced computational cost, making it ideal for real‑time research applications.
Google conducted side‑by‑side evaluations comparing NotebookLM on the old Gemini 3.1 branch with the updated 3.5 version. The company tested across five ''core evaluation dimensions'': Accuracy and Quality, Multilingual Support, Large Document Analysis, Document Creation, and Advanced Research. According to Google, the upgraded NotebookLM achieved a 65 percent win rate over the older model across these dimensions. The company did not release specific scorecards, but the consistent improvement suggests that users can expect noticeably better answers, especially when dealing with long documents or non‑English sources.
Perhaps the most intriguing addition is Antigravity, a cloud‑based code execution environment embedded directly in NotebookLM. The name is whimsical, but the capability is serious: Antigravity allows NotebookLM to write and run code as part of its research process. For example, a user could ask the AI to analyze a dataset, and NotebookLM could generate Python scripts, execute them in its own sandbox, and return visualizations or summary statistics. Google says NotebookLM will come with a library of more than 100 pre‑built software skills that can be combined into workflows, eliminating the need to switch between multiple applications. This effectively turns NotebookLM into a mini data‑science workstation that lives inside the note‑taking interface.
The update also expands NotebookLM's output capabilities beyond plain text and audio overviews. The new ''Studio Panel'' lets users generate a variety of file types, including data visualizations and charts (PNG, SVG), documents (PDF, DOCX, Markdown, text), images with Nano Banana (a lightweight image‑editing tool), structured data (CSV, JSON), Excel spreadsheets (XLSX), and PowerPoint presentations (PPTX). All generated files can be further refined through conversational prompts – for instance, asking the AI to change a chart's color scheme or convert a PDF to a slide deck. This positions NotebookLM as a one‑stop shop for research, analysis, and content creation.
Another significant improvement is in source discovery. Previously, NotebookLM required users to manually upload documents or paste URLs. Now, users can ask the AI to find relevant sources directly from the chat interface. The AI returns a ''research report'' listing potential sources, and the user can choose to import all or some of them. Once imported, those sources become part of the notebook's context for all future queries, enabling deeper, iterative research without leaving the app. This feature relies on Google's web search infrastructure and the same grounding technology that powers the company's AI Overviews.
The Gemini 3.5 Flash model itself is notable for its efficiency. At Google I/O, the company highlighted that the Flash model can reduce token costs by up to 80 percent compared to earlier models while maintaining quality. For enterprise users who process large volumes of content, this cost saving can be transformative. NotebookLM's integration with Flash means that heavy research sessions – analyzing hundreds of pages, generating multiple reports – will be less expensive to run, though Google has not yet announced specific pricing for the AI Ultra subscription that includes these features.
Behind the scenes, NotebookLM also benefits from improvements in multilingual support. The model can now comprehend and generate text in dozens of languages with greater accuracy, making it useful for international research teams. Its large‑document analysis capability has also been enhanced; the new system can handle documents that are tens of thousands of words long without losing context, thanks to Gemini 3.5's extended context window.
The rollout strategy follows Google's typical pattern: the update is first available to AI Ultra subscribers and Workspace business customers with AI Ultra Access or Expanded Access. These are the same tiers that already provide advanced features in Gemini for Workspace and Google Cloud. Standard consumer Google accounts will gain access in the coming weeks. Google hasn't specified exact dates, but the company has been accelerating the pace of feature releases to compete with offerings from OpenAI, Microsoft, and Anthropic.
NotebookLM's unique value proposition – source‑grounded research with flexible output – has made it a favorite among academics, journalists, and business analysts. The addition of code execution and file generation moves it from a passive note‑taking tool to an active productivity platform. Antigravity, in particular, opens up possibilities for data scientists and engineers who want to prototype analyses without leaving their browser. The 100+ pre‑built skills cover common tasks like data cleaning, statistical testing, natural language processing, and web scraping, reducing the need to write boilerplate code.
However, the update also raises questions about data privacy and security. By executing code in a cloud environment, users are essentially sending their data and code to Google's servers. For sensitive research – such as medical or legal documents – this may be a concern. Google has stated that all code execution takes place in a sandboxed environment and that data is handled according to its standard privacy policies, but the company has not provided detailed technical documentation about isolation methods. Users should evaluate their own compliance requirements before uploading confidential material.
Competitively, NotebookLM now overlaps with tools like Microsoft Copilot, which also integrates code interpretation and file generation, and with dedicated research assistants like Scispace or Elicit. But NotebookLM's tight integration with Google's ecosystem – including Drive, Search, and Vertex AI – gives it an advantage for users already invested in Google services. The new source‑discovery feature, in particular, leverages Google's search index in a way that competitors cannot easily replicate.
In closing, this update represents a maturation of NotebookLM from a simple experiment into a full‑fledged research platform. The combination of Gemini 3.5's efficiency, Antigravity's code capabilities, and the expanded file‑generation options makes it a compelling tool for anyone who needs to synthesize information from multiple sources. As the AI industry continues to evolve, NotebookLM is a clear signal that Google intends to stay competitive in the productivity space, not just by improving model quality but by reimagining the user interface for knowledge work.
Source:Ars Technica News