By Google rewired NotebookLM on June 8. The research tool that used to summarize your PDFs now writes Python, executes it in a cloud sandbox, and hands you an Excel file with the results. Backed by Gemini 3.5 and the Antigravity agentic framework, this is the most significant upgrade since the product launched — and it changes where NotebookLM code execution fits in a developer’s daily stack.
Code Execution Is Real — Not a Code Block You Copy
The headline feature is not “AI that suggests code.” Every notebook now gets a secure, isolated cloud computer that actually runs Python. You give NotebookLM a messy dataset or a complex question. It writes a script, executes it server-side, shows you the code, then shows you the output. The distinction matters: this is computation, not generation.
The system draws on over 100 curated software skills — pre-built capabilities covering data transformation, statistical analysis, chart generation, and file conversion. You don’t write the Python yourself. You describe what you need, Antigravity plans the steps, and the notebook executes them.
What Antigravity and Gemini 3.5 Actually Change
NotebookLM previously ran on Gemini 1.5 as a retrieval-augmented system: it answered questions by searching your uploaded documents. The new version is architecturally different. Gemini 3.5 provides the reasoning layer. Antigravity — Google’s agentic platform from Google I/O 2026 — provides the planning layer. The system can now chain tasks that previously required manual handoffs: find source, research gap, run calculation, produce report.
Source attribution stays intact. Every claim traces back to a specific document in your notebook. And for the first time, you can start a notebook with just a question — no pre-assembled source library required. NotebookLM will search the web, surface credible sources, and ask for your approval before proceeding.
The 11 Output Formats
This is the feature that closes the last mile. Previously, NotebookLM analysis lived inside the tool. Now you get downloadable, editable deliverables:
- PDF reports with embedded charts and tables
- Word documents (.docx)
- Excel spreadsheets (.xlsx) with computed tables
- PowerPoint presentations (.pptx)
- CSV and JSON structured data
- Markdown and plain text
- Data visualizations (PNG, SVG)
- AI-generated images via Nano Banana (PNG, JPG, GIF)
You can provide formatting instructions before generation and request edits after. For teams producing research-backed reports, the pipeline from “upload documents” to “download a client-ready PDF” now lives entirely inside NotebookLM.
MCP Integration: The Developer Angle
The feature that matters most to developers might not be in the official announcement. A community-built MCP server — notebooklm-mcp-cli — gives programmatic access to NotebookLM via Model Context Protocol. Query notebooks, add sources, and download outputs from your terminal, Claude Code, Antigravity, or any MCP-compatible agent.
# Install
uv tool install notebooklm-mcp-cli
# Configure for Antigravity: .gemini/antigravity/mcp_config.json
{
"mcpServers": {
"notebooklm-mcp": {
"command": "uvx",
"args": ["--from", "notebooklm-mcp-cli", "notebooklm-mcp"]
}
}
}A practical walkthrough on Dev.to covers the full setup with Gemini CLI and Antigravity. Google also publishes an official NotebookLM Enterprise API on Google Cloud for teams building at scale.
Who Actually Gets This
The new capabilities are exclusive to Google AI Ultra subscribers and Workspace accounts on AI Ultra Access or AI Expanded Access tiers. Google AI Ultra costs $249 per month. Free-tier users keep the old version — no NotebookLM code execution, no expanded output formats, no Gemini 3.5. Google says broader rollout is coming, but gave no timeline. Web only at launch; mobile is explicitly excluded.
What This Is (and What It Is Not)
NotebookLM does not replace Cursor or Claude Code. If you need to modify your codebase, build a feature, or debug production issues, you want a coding agent. NotebookLM’s strength is loading and reasoning over large corpora: API documentation, research papers, customer feedback datasets, transcripts.
The new version makes it genuinely useful for the step that precedes coding: understanding the domain, synthesizing research, and producing structured outputs that feed into your build environment. Pair it with Antigravity or a coding agent for the execution phase.
Google’s internal benchmarks claim a 65% win rate over the previous NotebookLM across five evaluation dimensions — 78.2% on web research tasks, 69.9% on large document analysis. Treat those as directional. They’re self-reported, the methodology is undisclosed, and no independent benchmarks are published yet.
The upgrade is real. The paywall is real. If you’re already on Google AI Ultra, test the code execution on a dataset you’ve been meaning to analyze. If not, the MCP integration is worth bookmarking for when access eventually reaches lower tiers — and check Google’s official announcement for when that happens.
