Four times faster than competing models, better at finding sources
Google has quietly transformed NotebookLM from a capable research assistant into something closer to an automated thinking partner, replacing its AI core with Gemini 3.5 Flash and weaving in coding tools that allow software tasks to run in parallel rather than in sequence. The upgrade reflects a broader ambition: that the distance between asking a question and acting on its answer should shrink to near nothing. For now, the most powerful features remain behind a subscription wall, a familiar tension between technological possibility and commercial reality.
- NotebookLM's leap to Gemini 3.5 Flash is not incremental — it processes dense, complex documents four times faster than rival models and outperforms Claude Opus on key benchmarks.
- The addition of Antigravity's parallel AI agents introduces a fundamentally different way of working: instead of one task at a time, multiple agents now execute simultaneously inside a live virtual machine.
- Over 100 pre-packaged software skills mean users can automate data consolidation, script generation, and research workflows without writing a single line of code themselves.
- New export formats — Word, PowerPoint, Excel, PDF, and JSON — close the gap between NotebookLM's output and the tools people actually use, reducing the friction of manual conversion.
- The most advanced capabilities are gated behind Google's AI Ultra subscription tier, leaving the broader user base waiting as Google signals a gradual, staged rollout ahead.
Google has overhauled NotebookLM, swapping its AI engine for Gemini 3.5 Flash and adding a suite of coding capabilities that push the tool well beyond its origins as a document summarizer. The platform has always occupied an unusual space — part note-taking app, part data analysis tool — built on the premise that AI can surface patterns humans might otherwise miss. This update deepens that premise considerably.
The move to Gemini 3.5 Flash brings a measurable performance jump. Google's own testing shows the new model handles research and source discovery better than its predecessor in roughly 78 percent of cases, with particular strength on lengthy, dense documents. The company also claims it generates responses four times faster than competing models and outperforms Claude Opus across several benchmarks — notable given that it's positioned as an entry-level offering.
The more significant shift is the integration of Antigravity, Google's code editor. Each NotebookLM workspace now includes a virtual machine where the tool can write and execute code, with Antigravity's parallel AI agents breaking projects into discrete tasks and running them simultaneously. A data scientist could ask the system to pull spreadsheets from multiple sources and consolidate them into a single format — and more than 100 pre-built software skills help users get reliable results from exactly these kinds of requests.
The update also expands output formats to include Word, PowerPoint, Excel, PDF, and JSON — the last of which allows NotebookLM's results to feed directly into other applications without manual conversion. For now, these capabilities are limited to Google AI Ultra subscribers and select Workspace add-on tiers, with broader availability and additional file format support planned for future updates.
Google has given NotebookLM a significant overhaul, swapping out the underlying artificial intelligence engine for Gemini 3.5, the company's newest family of large language models. The upgrade arrives alongside a suite of coding capabilities that aim to automate tasks developers and data scientists typically handle manually.
NotebookLM occupies an interesting middle ground in Google's product lineup. It functions as both a note-taking application and a data analysis platform. Users upload business documents and the tool summarizes them. Students collect research for assignments. Teams build presentations from source material. The service has always been built around the idea that AI can compress information and surface patterns humans might otherwise miss.
The move to Gemini 3.5 Flash represents a meaningful performance jump. Google's own testing shows the new version handles research and source discovery tasks better than its predecessor in roughly 78 percent of cases. It's particularly stronger when working with lengthy documents—the kind of dense material that often defeats simpler tools. The company also claims Gemini 3.5 Flash generates responses four times faster than competing large language models, and benchmarks suggest it outperforms Claude Opus across several measures, despite being positioned as an entry-level offering.
But the real shift in this update is the addition of coding features. NotebookLM now includes integration with Antigravity, Google's code editor that debuted last November. Antigravity's core innovation is its ability to break a software project into discrete tasks and assign each one to a separate AI agent. Running multiple agents in parallel is substantially faster than completing work sequentially. Each NotebookLM workspace now includes access to a virtual machine where the tool can write and execute code. A data scientist could, for instance, ask the system to write a script that pulls spreadsheets from multiple sources and consolidates them into a single standardized format. Google has packaged more than 100 pre-built "software skills" into the system to help users get better results from these requests.
The update also expands what NotebookLM can produce. The service now exports to Word, PowerPoint, Excel, and PDF documents—formats that weren't previously available. JSON support is new as well, which matters because it allows users to feed NotebookLM's output directly into other applications without manual conversion. The tool has also improved at identifying data sources relevant to specific research questions.
These capabilities are not yet available to everyone. Access is currently limited to subscribers of Google's AI Ultra plan and organizations using Google Workspace's AI Ultra Access and AI Expanded Access add-ons. Google has indicated it plans to broaden availability over time, and future updates will add support for additional file formats. The company is essentially rolling out a more powerful version of a tool that was already useful, while keeping the most advanced features behind a subscription tier for now.
Citações Notáveis
Gemini 3.5 Flash is positioned as an entry-level model, yet outperforms Claude Opus 4.7 across several benchmarks— Google
A Conversa do Hearth Outra perspectiva sobre a história
What makes Gemini 3.5 Flash different from what NotebookLM was using before?
It's faster and more capable. Google's testing shows it handles research tasks better in about 78 percent of cases, and it generates responses four times faster than competing models. For someone working with thick documents or trying to find relevant sources, that's a real difference.
The coding features sound like the bigger story here. What can someone actually do with them that they couldn't before?
They can ask NotebookLM to write and run code directly. Imagine a data scientist who needs to merge spreadsheets from five different sources into one clean dataset. Before, they'd write the script themselves or use a separate tool. Now they can describe what they need and the system handles it. The parallel agents part matters too—multiple tasks run at once instead of one after another.
Who gets to use this right now?
People with Google's AI Ultra subscription, and organizations paying for the AI Ultra add-ons in Google Workspace. It's not open to everyone yet. Google says they'll expand access over time.
Why lock it behind a subscription?
These are computationally expensive features. Running virtual machines and parallel AI agents costs more than basic summarization. Google is probably testing how people use them before making them widely available.
What's the practical impact for someone doing research or writing code?
Time savings, mainly. Better source discovery means less manual digging. Automated code generation means less time writing boilerplate. It's not revolutionary, but it's the kind of incremental improvement that compounds when you're doing this work every day.