Yao Shunyu, a legend from Tsinghua University, has made great contributions. The new Gemini has dominated programming overnight, and only 7 people in the world can beat it.

Just now, Google's Gemini 3 Deep Think has evolved on the spot! It ranked seventh globally in the Codeforces competition, defeating human players. In just three months, it has comprehensively shattered the State-of-the-Art (SOTA) records.

Unexpectedly, Google DeepMind has pulled out another big move late at night!

Today, Gemini 3 Deep Think has received a significant upgrade, almost shattering the SOTA records across all fields, marking that AI reasoning ability has entered a new dimension.

Yao Shunyu, a Chinese scholar who left Anthropic and joined Google, participated in the development of Gemini 3 Deep Think.

This time, in the fields of scientific research and hardcore engineering, Deep Think can be regarded as a "super brain".

It can render a sketch into a high-fidelity and practical 3D diagram of a laptop stand and directly print it out.

Google's VP showed off this side project, and this is what the final product looks like.

How powerful is the new version of Deep Think?

In the programming world, it scored an amazing 3455 Elo points, reaching the level of a world champion and ranking among the top 10 human players in the Codeforces competition!

That is to say, only 7 people in the world have defeated Gemini 3 Deep Think. A year ago, the strongest o3 only scored 2727 Elo points.

In the Human Last Exam (HLE), Gemini 3 Deep Think refreshed the SOTA record, achieving a score of 48.4%.

Even more, it saturated the most difficult ARC-AGI-2 benchmark overnight, leading the way with a new SOTA score of 84.6%.

Compared with Gemini 3 Pro, Deep Think has comprehensively outperformed it and left Claude Opus 4.6 and GTP-5.2 far behind.

What's even more amazing is the real - world performance of Gemini 3 Deep Think.

Lisa Carbone, a mathematician at Rutgers University, asked it to review a profound physics and mathematics paper during her research.

As a result, Deep Think actually discovered a subtle logical flaw that even human peer - reviewers had missed.

This epic evolution of Gemini 3 Deep Think has kept some people awake at night.

Currently, Google AI Ultra subscribers can experience the new version of Deep Think in Gemini. At the same time, it is also open to some researchers, engineers, and enterprises through the API for the first time.

Powerful enough to shatter SOTA, a grand slam of Olympic gold medals

Last year, the special version of Deep Think successfully solved the most difficult tasks in the field of reasoning and won gold medals in the world championships of mathematics and programming.

Just yesterday, Google DeepMind also did a pre - warm - up.

The "AI mathematician" Aletheia, backed by the first - generation Deep Think, can independently write papers and prove multiple difficult problems in the "Erdős conjecture".

Moreover, Deep Think directly overturned a decade - long conjecture and overcame 18 major research bottlenecks at once.

Now, the iterated Deep Think has refreshed the SOTA records in multiple high - difficulty benchmark tests:

• Human Last Exam (HLE): Set a new benchmark, with an accuracy rate of 48.4% without using tools;
• ARC - AGI - 2: Reached an unprecedented 84.6% and was verified by the ARC Awards Foundation;
• Algorithm competition platform Codeforces: The Elo rating reached an amazing 3455 points;
• 2025 IMO: Reached the gold - medal level.

In addition to mathematics and algorithm competitions, Gemini 3 Deep Think now also performs excellently in a wide range of scientific fields such as chemistry and physics.

In the written - test parts of the 2025 International Physics Olympiad and the International Chemistry Olympiad, the new version of Deep Think also has the strength to win a gold medal.

Moreover, it is also proficient in advanced theoretical physics, achieving a score of 50.5% in the CMT - Benchmark test.

Additionally, Gemini 3 Deep Think has reached the limit on ARC - AGI - 1.

In the official demonstration, Gemini 3 Deep Think can create a visualization scheme for the "Spatio - Temporal Cyclic Video Transformer" architecture based on a paper.

Deep Think enters scientific research, accelerating tenfold

In addition to its top - notch performance, Deep Think has also broken through the boundaries of intelligence and can solve modern problems in the fields of science, research, and engineering.

It can not only help researchers interpret complex data but also assist engineers in modeling physical systems through code.