Google releases Gemini 3 Deep Think

A reasoning mode for science and engineering. Found an error reviewers missed and broke olympiad records.

Author: Michael Kokin ·

Google released Gemini 3 Deep Think. This is a specialized reasoning mode designed for science, engineering, and complex research. What makes it good?

It catches what humans miss. A mathematician from Rutgers University ran a technical paper through the model and Gemini found an error that even peer reviewers had missed.
It turns sketches into 3D models. Draw something freehand — and the model generates files for 3D printing. Literally a bridge from imagination to the real world.
It doesn't fear chaos. The model was specifically trained to work with "dirty" data where there are no clear answers or variables contradict each other.
It breaks records. Gold at physics and chemistry olympiads, plus the best result on Humanity's Last Exam — a test specifically designed to stump AI.

The model uses "advanced parallel reasoning." Essentially, it simultaneously works through multiple hypotheses and combines them into one coherent solution.

For those working with code, complex data, or systems modeling — this is a must-have.

Access has been opened for Ultra subscribers in the app, and authorized researchers and developers can connect via API.