NotebookLM upgrades to Gemini 3.5, Antigravity, and new file types

Google’s NotebookLM just got one of its biggest upgrades yet. The product is moving to the latest Gemini 3.5 model, adding support for more file types, and tightening how it integrates web sources. Google also says Antigravity is now embedded, which it claims lets NotebookLM “do more” with all those queries. For anyone running AI workflows inside Google’s ecosystem, this is a very practical change: it is not just a model swap, it is an expansion of what your inputs can be and how reliably your system can turn sources into outputs.

The immediate consequence is that this upgrade is not universal. The change is only available for AI Ultra and enterprise accounts right now, which means most users will not see it immediately even if they are watching every AI announcement like it is stock-market news. Still, Google is giving enough detail to make the direction clear: NotebookLM is being brought in line with Google’s newer generation of Gemini performance, including what Google says are cost and speed benefits from Gemini 3.5 Flash. That last part matters, because teams do not adopt LLM tools just to be impressed. They adopt them to ship better internal answers while keeping token costs under control.

To understand why this matters beyond a product update, it helps to look at how NotebookLM is positioned. NotebookLM, which launched in 2023 at the beginning of the AI boom, is designed to analyze specific sources like documents and webpages using Google’s latest AI models. In other words, it is an interface for source-based work, not a general chatbot you can ask anything. That distinction is key for executives and operators because source grounding is where accuracy claims either hold up in practice or collapse under real-world use.

Google is positioning the upgrade as an objective improvement. It ran side-by-side evaluations of NotebookLM on the old Gemini 3.1 branch and the updated 3.5 version. Google is vague about the exact tests, but it still breaks results into “top five core evaluation dimensions.” Those dimensions are Accuracy and Quality, Multilingual Support, Large Document Analysis, Document Creation, and Advanced Research. In those evaluations, Google says NotebookLM averaged a 65 percent win rate versus the older model.

That 65 percent win rate is the kind of number that boards and procurement teams can use, because it suggests a consistent lift across multiple areas rather than a single headline feature. At the same time, Google’s category-level framing is a reminder that “LLM quality” is rarely one metric. For enterprise buyers, the question is usually what improves for the workflows that matter: large doc handling for compliance-heavy teams, multilingual output for global operations, and research and document creation for policy, marketing, and legal adjacent work. Google is effectively telling decision-makers that 3.5 is stronger across the full notebook-to-output chain.

The engine behind this is Gemini 3.5 Flash, which debuted at Google I/O earlier this year. Google says Flash is faster and more efficient, and it claims companies that worry about token costs can save big by moving projects to the new Flash model, while also getting outputs of similar or better quality. In this update, the improvement story is meant to travel downstream: NotebookLM is getting the benefit of the same performance direction, even though the headline upgrade emphasizes Gemini 3.5 and not Flash specifically. Still, for teams measuring unit economics of AI, the practical takeaway is that Google is trying to make higher-quality outputs more affordable.

There is also a strategic subtext here about AI governance and rollout control. Limiting the new features to AI Ultra and enterprise accounts gives Google a way to manage load, monitor performance, and keep tighter quality guarantees while it scales. It also creates a tiering incentive, where higher-paying customers can adopt new model capabilities earlier. For decision-makers, that means vendor evaluation timing matters. If you are planning an internal launch or renewing contracts, “when you get the upgrade” can be as important as “what the upgrade is.”

If you are a peer operator, founder, or investor evaluating the competitive landscape, this is a signal that source-based AI products are moving from novelty to infrastructure. NotebookLM is being treated like a system that needs better models, better input coverage (more file types), and smoother web source integration. Antigravity is the most opaque word in the announcement, but Google is clearly framing it as functional, embedded support that increases what NotebookLM can do with queries. Combine that with the side-by-side evaluation framing and the token cost narrative from Gemini 3.5 Flash, and you get a coherent story: Google wants NotebookLM to deliver more value per query, at scale, and within enterprise guardrails.

For enterprises deciding what to deploy now versus what to pilot later, the stakes are straightforward. If NotebookLM quality meaningfully improves, it reduces rework and increases trust in generated outputs. If costs can be kept in check by shifting to more efficient models, it reduces the risk that AI becomes a line-item whose size shocks finance. And because this upgrade is gated today to AI Ultra and enterprise accounts, the smart move is to plan around rollout windows, not just roadmap wishlists.