NotebookLM Major Upgrade: Supports Antigravity to Help You Write Code and Run Analysis, Switched to Gemini 3.5 Flash Engine

Google has swapped out the engine in NotebookLM for the Gemini 3.5 Flash engine, and it comes with an Antigravity agent-style execution environment built in. This upgrades the research tool from a “passive Q&A notebook” to an AI research assistant that “can write code itself, run analyses, and complete workflows end to end.”
(Background: After community backlash, Google admits the mistake—Antigravity Gemini rate limits increased by 3x, and this week’s quota is reset.)
(Additional context: Google Gemini API hit a “cache billing vulnerability,” and developers deleted invalid data and got hit with heavy charges of 20,000 reals.)

Table of Contents

Toggle

• What changed inside the notebook
• Antigravity gives the notebook hands and feet
• Who gets it first, and what it means

Google has a well-known habit: killing products. Reader was axed, Stadia was shut down, Inbox was scrapped—this list is so long that someone even built a website to document it. But NotebookLM, born in 2023 as one of the company’s earliest generative AI products, didn’t just survive—it has also received the biggest upgrade of its kind to date, right now.

This upgrade isn’t just a matter of swapping out the engine. Google has pushed “agent-style execution capability” directly into a consumer research tool, turning the notebook from “ask it and it answers” into “it can write code by itself, find information, and run analyses.”

What changed inside the notebook

The core upgrade this time for NotebookLM is switching to Gemini 3.5 Flash. This model was officially unveiled at Google I/O 2026 this year. Google claims (in its official wording) that it is 4 times faster than other leading models, beats Gemini 3.1 Pro across nearly all benchmarks, and that enterprises migrating projects to Flash can clearly reduce token costs.

In real-world tests, Google’s official announcement states that it conducted side-by-side evaluations of the two NotebookLM versions (Gemini 3.5 vs the 3.1 branch): across five dimensions—“accuracy and quality,” “multilingual support,” “large file analysis,” “file building,” and “advanced research”—the Gemini 3.5 version averaged a 65% win rate (Google’s self-assessment).

However, Google’s explanation of how it designed and conducted these evaluation dimensions and methodology is quite vague, and outsiders cannot independently verify this number—so it should be kept at arm’s length.

Beyond the engine swap, NotebookLM also adds support for more file formats and simplifies the process of integrating online sources. Users can, even when they only have “loose ideas and questions,” have NotebookLM automatically search Google to find high-quality online sources and build a library of data sources.

Antigravity gives the notebook hands and feet

The most critical step in this upgrade is that NotebookLM now includes support for Antigravity. Antigravity is Google’s “agent-first” AI development platform introduced at I/O 2026. Version 2.0 already covers desktop apps, CLI, and the SDK. Multiple AI agents can automatically plan tasks, write code, open browser tests, fix bugs, and deploy—allowing the entire software development cycle to be driven by AI.

Worth noting: Google has integrated the previously standalone Gemini CLI and turned it into the Antigravity CLI. This means Google is converging and unifying its fragmented lineup of agent-style tools into a single stream. Antigravity isn’t just an embedded feature—it’s the consolidation point for the entire agent-based development ecosystem.

Now, this Antigravity capability is being introduced into NotebookLM. Each notebook comes with a “secure cloud computer,” enabling NotebookLM to write and execute code directly within your research context via Antigravity—not just “give you a snippet of code and let you run it yourself.”

This represents a clear shift in direction. The concept of agentic AI has already appeared for a while in products from vendors like OpenAI and Anthropic. Google is now pushing this capability into broader everyday use cases through NotebookLM—the target isn’t engineers, but researchers and knowledge workers.

Who gets it first, and what it means

This upgrade is initially available to Google AI Ultra subscription users, as well as Workspace business customers who have AI Ultra privileges; users on other plans will need to wait for subsequent releases.

For people doing research work, the more practical question is: how stable are these features in real use? What kinds of mistakes will Antigravity’s agent-style execution make on complex tasks? As NotebookLM expands into “active execution,” the possibilities and ways it can go wrong also become more diverse. Only after it’s actually released can we know.

A tool that once merely “let you ask questions about files” now claims it can write code for you, run analyses, and generate presentations. Whether this deserves serious attention depends on whether Google can make Antigravity’s agent-style capabilities truly reliable in research scenarios that aren’t engineering-focused—not just look impressive at launch events.