Google upgrades NotebookLM with Gemini 3.5 and Antigravity

The headline change is not the model name—it is that every notebook now boots its own secure cloud computer, a virtual machine NotebookLM can use to write and execute code on your behalf ^[1]. Under the hood this runs on Google's Antigravity coding model alongside Gemini 3.5, and the sandbox ships with more than 100 curated 'software skills'—pre-packaged capabilities the assistant can call to do real work rather than just describe it ^[2].

In practice that means NotebookLM is no longer limited to summarizing what you give it. Ask it to crunch a messy dataset and it can write a script, run it in the sandbox, and hand back an actual spreadsheet or chart instead of a paragraph describing one. Google also says the upgrade exposes more of the model's reasoning, so you can see how it reached a conclusion rather than trusting a black box ^[1]. The architectural leap is from a language model that talks about data to an agent that manipulates it.

For two years NotebookLM has been a reading assistant: you fed it documents, it answered questions grounded in them. This update flips the starting point. You can now open a notebook with nothing but a rough question, and NotebookLM will use Google Search to find relevant, high-quality sources and build the library for you—while still letting you approve what gets added ^[3].

That reframes the product as a research agent that plans and executes multi-step work rather than a passive question-and-answer surface, and it puts NotebookLM into direct competition with conversational research tools like ChatGPT and Claude ^[2]. The bet is that source-grounded answers plus autonomous source-finding beat either capability alone. Independent coverage frames it the same way—as a shift that 'reduces the need for switching between multiple applications, streamlining processes while minimizing errors' ^[4]—though that very convenience is what raises the stakes on accuracy.

Google is leaning hard on benchmark numbers to sell the upgrade. In its own side-by-side evaluations, the new NotebookLM posted an average win rate above 65% across five core categories—a 15-point margin over parity—climbing to 69.9% on large-document analysis and 78.2% on advanced web research and source discovery ^[1]. Google further claims the system generates output roughly four times faster than competing large language models, and positions the underlying Gemini 3.5 Flash as beating Anthropic's Claude Opus 4.7 on several benchmarks ^[2].

The caveat worth holding onto: these are Google's own internal comparisons against the previous version of NotebookLM, not a neutral third-party benchmark. A 65% win rate means the new version was judged better most of the time—not that it is right 65% of the time. The numbers are a credible signal that the reasoning engine improved meaningfully, but they measure preference in a controlled evaluation, not real-world reliability.

The upgrade is not flipping on for everyone. At launch it is gated to Google AI Ultra subscribers and a slice of Workspace business customers on the AI Ultra Access and AI Expanded Access add-ons, with wider availability promised later ^[2]. So the most capable version of NotebookLM is, for now, a premium tier—a notable choice for a product that built its audience as a free tool.

There are technical limits too. Even sympathetic coverage notes the assistant still 'struggles with more complex coding tasks and nuanced technical explanations' relative to dedicated coding tools ^[4], and analysts caution that a smoother, more finished-feeling workflow can invite overtrust—users accepting polished output without checking it against the underlying sources ^[4]. The community reaction to NotebookLM's broader move onto Gemini has carried the same ambivalence: enthusiasm for sharply better reasoning, paired with friction over staged access and pointed complaints from heavy users about how quickly the new, more compute-hungry experience burns through usage limits. The capability jump is real; so is the bill that comes with running a code-executing agent for every notebook.