Google dropped a major upgrade to NotebookLM on June 8, 2026, transforming it from a smart document reader into a full research agent. Rebuilt on Gemini 3.5 and Antigravity (Google’s agent-first coding IDE), it now gives every notebook its own secure cloud computer.
Previously, it required manual uploads, offered no code execution, hid its reasoning, and had limited outputs. Now it supports chat-driven source discovery via Google Search, shows step-by-step thinking, runs real code with over 100 built-in skills, and exports to more than a dozen formats including native PPTX, XLSX, DOCX, PDF, CSV, and SVG. Best of all, you can start from a completely blank notebook.
The standout feature is the per-notebook secure cloud computer. NotebookLM can write and execute code against your sources to clean messy datasets, normalize dates and currencies, run accurate math and stats, generate charts, and assemble professional outputs. It feels like having a junior analyst with a sandboxed laptop.
Chat-driven source discovery is equally powerful. Describe your project in a blank notebook, and it uses Google Search to suggest high-quality sources—including foreign-language primary materials and related works by authors. You review and curate what to keep instead of having to upload everything upfront.
Output capabilities are enhanced as well. You can request editable PowerPoint decks, functional Excel spreadsheets, polished PDFs, DOCX reports, data files, and images, then iterate with follow-up instructions.
These upgrades unlock strong new workflows: zero-upload literature reviews with citation matrices, turning dirty data into finished PDF reports, and generating board-ready competitive briefs from scratch. It can even export raw CSV/JSON for verification.
Google reports solid gains: 65%+ average win rate, ~70% on large document analysis, and 78%+ on web research and source discovery. These are internal benchmarks, but they match the new agent strengths.
The access is still limited. Only Google AI Ultra and Workspace AI Ultra users will have the opportunity to use this on the web for now. Auto-discovered sources need careful vetting for quality and bias. Foreign-language material requires extra scrutiny, and data stays in Google’s cloud.
Overall, NotebookLM has taken a huge leap toward true agentic research. It’s not perfect yet, but the combination of discovery, code execution, and professional outputs makes it far more capable.
Who’s already using the new version? What’s your first project—research, data work, or presentations?
TL;DR: NotebookLM now starts from a blank page, finds sources with Google Search, analyzes with real code, shows its thinking, and delivers editable PPTX, XLSX, PDF and more. A major evolution from simple summaries.
