NotebookLM Gets Gemini 3.5 Flash, Beats Old Model by 65%

NotebookLM now runs on Gemini 3.5 Flash, and Google’s internal side-by-side evaluations show a 65% win rate against the old Gemini 3.1 branch across five core dimensions.

What Changed: The 3.5 Flash Upgrade

Google shipped NotebookLM’s biggest update since launch: the underlying model jumps from Gemini 3.1 to Gemini 3.5 Flash, the same architecture unveiled at Google I/O this year. The company claims Flash delivers much faster and more efficient processing while cutting token costs — a direct pitch to organizations that hit budget ceilings with older models. In practice, that means NotebookLM can now chew through larger documents and handle multilingual queries without the latency or cost wall.

Google evaluated NotebookLM across five “core evaluation dimensions”: Accuracy and Quality, Multilingual Support, Large Document Analysis, Document Creation, and Advanced Research. Across the board, the 3.5-based version won 65% of head-to-head comparisons. No breakdown per dimension is provided, but a two-thirds sweep is hard to ignore.

Antigravity and the Deeper Query Layer

Alongside the model swap, NotebookLM gains embedded support for something Google calls “Antigravity.” The source material doesn’t detail what Antigravity is — likely an internal retrieval or context-expansion tool — but the description says it lets the notebook “do more with all those queries.” Combined with expanded file-type support and streamlined web source integration, NotebookLM moves from a simple document Q&A tool toward a multi-source research assistant that can weave in live web content.

Why This Matters Beyond NotebookLM

This isn’t just a refresh of a single product. Google is proving it can migrate its consumer AI tools onto the Flash model without regressing quality — and at a 65% win rate, it improved. For teams building on Google’s AI stack or evaluating cost-performance tradeoffs, the Flash model now has a real deployment in a high-profile app. If NotebookLM’s 65% advantage holds in third-party benchmarks, the old Gemini 3.1 branch becomes dead weight.

The next question is whether Antigravity and the new file-handling pipeline get extracted into a standard API. If so, developers get a ready-made retrieval-augmented generation stack that already works at Google scale.

Source: Gemini 3.5 and Antigravity come to Google NotebookLM
Domain: arstechnica.com