Google just flipped a huge switch inside NotebookLM. The popular research assistant just gained the ability to write and run actual Python code directly in the browser. It shifted from a passive reading tool into an active data-computing engine almost overnight. I was honestly skeptical at first. But a recent hands-on test proves the June 8 platform upgrade is pulling off things I genuinely didn’t expect.
During a hands-on review conducted by Mahnoor Faisal at XDA, NotebookLM tore through a disorganized academic grading spreadsheet and complex professional traffic data. The tool didn’t just crunch the numbers. It independently identified a hidden formula error in the grading sheet that shorted a student 14 marks. It also caught mathematical inconsistencies in device traffic shares without ever being prompted to run an audit. The AI actually “thinks out loud” now. It details its data-cleaning methodology step-by-step before generating clean, typeset deliverables like PDF reports and histograms right in its Studio panel.
What We Know
- The code execution runs in a secure cloud virtual machine attached specifically to your notebook.
- It uses Python as its primary scripting language for data analysis.
- The system generates visual outputs directly inside the interface, removing the need to export raw data to Excel.
- This marks the biggest functional jump for the tool since its original Google Labs debut three years ago.
This specific capability is all thanks to a major architectural overhaul. NotebookLM now runs entirely on Gemini 3.5 and Antigravity. The update provides each individual notebook with its own isolated, sandboxed cloud computer. On top of that, the system is loaded with over 100 pre-built software skills it can reach for mid-task to solve whatever problem you throw at it.
The performance bump here is wild. According to Google’s internal metrics, the new system hit an average win rate of 65% against its prior iteration. And when you look at the underlying Gemini 3.5 Flash model, it is absolutely crushing previous versions. It scored 76.2% on the Terminal-Bench 2.1 coding benchmark and an impressive 83.6% on MCP Atlas, easily outpacing Gemini 3.1 Pro across the board.
How Gemini 3.5 Disrupts the Data Analysis Pipeline
We are looking at a very different kind of AI interaction here. Until now, Large Language Models have been notoriously bad at math. They hallucinate numbers because they predict text, not equations. By giving NotebookLM a secure cloud computer to physically run Python scripts, Google is bypassing the LLM math problem entirely. The AI writes the code, but the virtual machine does the actual calculation.
This directly challenges dedicated data-science copilots. If a free research tool can automatically clean a messy CSV file, catch a human formula error, and instantly spit out a typeset PDF histogram, everyday users no longer need to learn Pandas or SQL for basic data crunching. It democratizes complex data analysis in a way that standard chatbots simply can’t match without native execution environments.