Isomorphic Labs Uses Gemini 4 Argon to Parse 500-Page Patents in a Day Instead of a Month

Isomorphic Labs has cut the audit of chemical patents of up to 500 pages from about a month to one day. Company researchers Brian Bender and Santosh Adhikari use the multimodal Gemini 4 Argon model for the task. It reads dense scientific papers and patent texts and builds structured reports from them. This comes from a Google DeepMind video published on 13 October 2026.

Where the time goes

Chemical patents are written to protect as much as possible, not to be easy to read. A molecule's structure may appear on one page, while the experimental data for it, such as activity measurements, sits on another, 50 pages away from the first. A person checking the literature has to hold dozens of such links in their head and keep flipping back and forth through the document. The result is weeks of work and errors that are hard to spot: a mismatched table or a missed example.

Isomorphic Labs, an Alphabet subsidiary that grew out of DeepMind, develops drugs with the help of AI. For the company, analyzing other people's patents and publications is part of the earliest stage of research. That is where the bottleneck appears, according to the video's authors: before a team starts designing molecules, someone has to understand what is already known and patented in the field.

What the model does

Bender and Adhikari use Gemini 4 Argon as a synthesis tool. The model receives the entire document, including images of chemical structures, not just the text. It then extracts data from the patent and arranges it into a structured report. According to DeepMind's description, this takes one day instead of a month, and teams find interrelated facts faster that are easy to lose in manual reading.

Multimodality solves a specific problem here. If the model sees only extracted text, a molecule's formula becomes a gap or a garbage string. If it works with an image of the page, it can link the drawn structure to a results table in another part of the document.

What the material tells us:

• model: Gemini 4 Argon, multimodal;
• task: auditing chemical patents and scientific papers;
• document size: up to 500 pages;
• gap between structure and data: about 50 pages;
• output: a structured report;
• turnaround: one day instead of a month;
• users: Isomorphic Labs researchers Brian Bender and Santosh Adhikari.

What the video leaves out

This is promotional material, and its figures come without a methodology. DeepMind does not say how extraction accuracy was measured, what share of structures and values the model recognized correctly, how much time a human spent checking the report, or which patents the model was tested on. In addition, "a month versus one day" compares a full manual audit with the model's first pass. If an expert still double-checks the result, the real savings will be smaller. The video also does not give Gemini 4 Argon's parameters, context window or price; DeepMind points to the Google blog for details.

What teams working with patents should check

If you have a similar task, start not with a demo but with your own benchmark. Take 5–10 documents that your specialists have already analyzed by hand and compare them with the model's reports on three measures: structures found, data correctly matched, and omissions. Measure separately how long a human takes to check the report, because that determines whether the multiple-fold gain survives.

It is also worth deciding in advance which errors are acceptable. A missed example in a literature review is annoying but not critical. A wrong link between a molecule and its activity can distort a decision about further experiments. For that reason, it is best to build the structured report with references to source pages: the reviewer should be able to jump from a line in the report to the original fragment of the patent in seconds.