What the Gemini 3.5 Upgrade Means for Everyday Research
NotebookLM’s latest upgrade is an AI research workspace that combines Gemini 3.5 reasoning, cloud-based code execution, and agentic research capabilities to help users discover sources, analyze data, and produce ready-to-share outputs in one place. Instead of jumping between search, notebooks, and separate data analysis AI tools, students, analysts, and knowledge workers can now start with a question and grow a complete research project inside a single notebook. The chat experience now runs on Gemini 3.5 and Google’s Antigravity technology, which boosts accuracy and consistency while exposing more of the AI’s step‑by‑step thinking in the conversation. Google reports that the upgraded NotebookLM achieves an average win rate of more than 65 percent across five core evaluation categories compared with the previous version, including strong gains in large‑document analysis and advanced web research.
NotebookLM Code Execution Turns Chat into a Data Lab
One of the most significant changes is that every notebook now includes a secure cloud computer, turning NotebookLM into a lightweight analysis environment. Within a chat, the AI can write and run code on the user’s behalf, drawing on more than 100 curated software skills to work across uploaded sources. For students, that can mean running quick statistical checks on study data. For analysts, it enables building custom models, cleaning CSV files, or generating data visualizations without leaving the NotebookLM interface. This cloud-based NotebookLM code execution is aimed at deeper, repeatable workflows: the model can run scripts, refine them, and then summarize what the results mean in plain language. Because the code runs in a secure cloud computer tied to each notebook, users keep their analysis context together with the documents, transcripts, and notes that feed it.
Agentic Research Capabilities Automate Source Discovery and Context
Previously, NotebookLM was strongest once users had already collected PDFs, web pages, and documents. The new agentic research capabilities shift that starting point. Users can now open a blank notebook, ask a broad question, and let NotebookLM propose and gather relevant materials. The system can build a source repository directly in chat, use Google Search to discover related content, and identify primary sources in different languages, while still keeping users in control of what is added. According to Google, the upgraded system reaches a 78.2 percent win rate in advanced web research and source discovery compared with the prior baseline, highlighting how agentic features expand its usefulness. Crucially, NotebookLM maintains attribution for each source, so users can see where a claim came from, inspect quoted passages, and trace reasoning back to original documents.
From Raw Inputs to Downloadable Reports, Charts, and Slides
Once the research and analysis are complete, NotebookLM now focuses on generating finished work products. The tool can turn a notebook’s contents into PDF reports with charts and tables, detailed budget-style summaries, or structured datasets ready for further processing. Users can export reports, charts, documents, spreadsheets, slide decks, and images in formats such as PDF, DOCX, XLSX, PPTX, CSV, JSON, Markdown, PNG, SVG, JPG, GIF, and plain text. You can give detailed export instructions—like specifying sections, tone, or level of technical depth—and then refine the file after it is generated. This closes a common gap in research workflows: instead of copying insights out of chat, users can move smoothly from question, to analysis, to a downloadable artifact that is ready for sharing with classmates, stakeholders, or colleagues.
How Expanded Reasoning Changes Multi‑Step Research Workflows
The deeper impact of these AI research automation tools shows up in long, multi‑step projects. NotebookLM can now read large document sets, identify themes, propose follow‑up questions, search for missing context online, run code to test hypotheses, and then package findings—all from the same chat. Google’s internal evaluations show a 69.9 percent win rate in large‑document analysis, suggesting that Gemini 3.5 features are better at holding complex threads over time. For students, that could look like turning a semester’s readings into a literature map and draft outline. For analysts and knowledge workers, it can mean moving from raw logs or survey data to consistent, shareable reports. Because users approve sources and can see reasoning steps, the system aims to support critical thinking instead of replacing it, making NotebookLM a flexible partner rather than a black box.
