Google Gemini 3 Deep Think can solve messy engineering problems
Google has rolled out a big update to its Gemini 3 Deep Think model, aiming to help solve tough science and engineering problems—even when the data is messy.
The new version is now live for Google AI Ultra subscribers on the Gemini app, and researchers can try it early through the API.
It set new records in various Olympiads
The upgraded model set fresh benchmarks: it scored 84.6% on the ARC-AGI-2 reasoning test, hit an Elo rating of 3,455 on Codeforces, and set new records on several AI tests.
It also did great in chemistry and physics Olympiads.
Gemini 3 Deep Think outperformed GPT-5.2, Claude Opus 4.6
Gemini 3 Deep Think beat top rivals like OpenAI's GPT-5.2 and Anthropic's Claude Opus 4.6 on key tests—including ARC-AGI-2 and Humanity's Last Exam—outperforming those rivals on those benchmarks.
It can even admit mistakes and learn from them
Researchers are putting it to work: Rutgers mathematician Lisa Carbone used it to spot errors in published math paper.
It can even turn sketches into files for 3D printing or power advanced math tools that browse the web and admit mistakes—pretty handy for anyone tackling tough projects or studies.