Yao Shunyu’s Google Debut: New Gemini Model Shatters SOTA Records, Only 7 Humans Defending Carbon

Source Domain: eu.36kr.com

Google launched Gemini 3 Deep Think to counteract advances by competitors like Claude Opus 4.6 and GPT Codex 5.3.
On Codeforces, Gemini 3 Deep Think achieved a prestigious Elo score of 3455, ranking 8th globally.
It surpassed the previous top Elo score of 2727 held by o3; also, it set a record of 84.6% on ARC-AGI-2, significantly higher than previous models.
In the Humanity’s Last Exam (HLE), Gemini 3 Deep Think achieved a 48.4% score, which is state-of-the-art (SOTA).
The new version of Deep Think is designed to excel in scientific research and engineering, with capabilities including analyzing sketches, modeling shapes, and generating 3D printing files.
Notable achievements include successfully identifying a logical flaw in a specialized mathematical paper that eluded previous manual reviews and optimizing processes for new semiconductor material growth.
The upgrade notably reduced reasoning costs from $77.16 per task to $13.62.
Gemini 3 Deep Think has won numerous SOTAs and gold medals in difficult benchmarks like the ARC-AGI series and International Mathematical Olympiad while also performing well in physics and chemistry Olympiads.
The development team includes prominent Chinese researchers, such as Yi Tay and Shunyu Yao, both of whom have significant contributions to the AI and physics fields.

You may have missed