Google's IMO Gold Medal-level Gemini 3 Launched Late at Night, Led by a Chinese Prodigy, OpenAI Powerless to Counter

It's so explosive!

Within just half a month, Google DeepMind has finally released the most powerful IMO gold - medal model, Gemini 3 Deep Think.

This summer, Gemini 2.5 Deep Think won gold medals in the IMO and ICPC international competitions respectively.

This time, Google has injected new blood into it - Gemini 3.

With its "parallel thinking" ability, Gemini 3 Deep Think can solve extremely difficult mathematical and scientific problems!

In the benchmark tests, Deep Think completely outperformed Gemini 3 Pro. Especially in the HLE, it scored 41% without using tools.

Meanwhile, on ARC - AGI - 2, it led the world with a score of 45.1%.

In the following example, with the same instruction, Gemini 3 Pro and the Deep Think version were asked to create an accurate interactive 3D scene based on a sketch of a museum pavilion roof.

Obviously, the latter has almost a 1:1 reproduction of the original image in terms of restoration, and in terms of interaction, the light and shadow changes conform to physical logic.

Today, Gemini 3 Deep Think has been launched on the Gemini App, and all Ultra users can experience it.

The most powerful IMO gold - medal model is here

Gemini 3 Deep Think has officially opened a new era of "deep thinking" and expanded the boundaries of intelligence once again.

Gemini 3 Deep Think is an iteration based on the previous generation, Gemini 2.5 Deep Think, and has achieved a qualitative leap in reasoning ability.

It is specifically designed to tackle complex mathematical, scientific, and logical problems that even today's top - tier models find difficult.

In multiple benchmark tests, Gemini 3 Deep Think has achieved industry - leading results.

It scored 41% in the Humanity’s Last Exam (without tool assistance) and set a new record of 45.1% in ARC - AGI - 2 (with code execution).

On ARC - AGI - 1 and ARC - AGI - 2, Gemini 3 Deep Think is unparalleled in strength.

The reason why it is so powerful is that Deep Think uses "parallel reasoning" to think simultaneously and can explore multiple hypotheses at the same time.

In more demos released by Google DeepMind, the power of Gemini 3 Deep Think can be seen.

As shown below, it was asked to create a 3D version of the domino game. In the level settings, Deep Think combines creativity and surprise, and also simulates real collision physics effects.

In this example, Gemini 3 Pro and Deep Think were respectively required to create a procedurally generated Earth - like planet in a single HTML file.

A side - by - side comparison clearly shows that Deep Think is more creative.

Some developers have been amazed by the actual tests and said that Gemini 3 Deep Think has amazing capabilities.

A photo of a glass bottle was uploaded, and it was asked to generate a 3D animation of a soft rubber version of the same bottle falling on the ground.

Some netizens have excitedly said that at this rate, we will truly achieve AGI in 2026!

Led by Chinese scientists, the AGI ace team is sprinting

Also today, Google DeepMind officially announced that it will form a new elite team in Singapore.

It will be led by Chinese scientist Yi Tay, specializing in advanced reasoning, LLM/RL, and promoting the development of the most cutting - edge SOTA models such as Gemini and Gemini Deep Think.

Yi Tay will report to the team led by Quoc Le (a Google Fellow - level expert) in Mountain View.

Coincidentally, this team is one of the core forces behind Gemini Deep Think's gold - medal wins in international competitions and has also made important progress in Gemini.

Yi Tay said that we will start with a small but extremely powerful team.

Because in the era of large models, "talent density" is more crucial than anything else.

Importantly, this team can also conduct joint research with legendary figures in the AI field, including "the King of Reasoning" Denny Zhou, the man behind "Banana" Mostafa Dehghani, and the "GOAT" of the AI world, Noam Shazeer.

Meanwhile, many "super teammates" have also joined the research, including Vinh Q. Tran, the inventor of generative retrieval, Thang Luong, the general in - charge of the IMO gold - medal project, Xuezhi, the pioneer of chain - of - thought, and Shane Gu, a top Japanese AI expert.

In the past few months, Google DeepMind has secretly recruited top - tier talents from around the world.

This project has received strong support from internal senior management such as Demis Hassabis and Jeff Dean.

Now, they continue to issue recruitment notices, and there is an opportunity to stand on the key track to AGI.

Before long, this Singapore team will soon become a team with extremely high combat effectiveness.

Gemini 3 surges with 15% traffic

With the strong release of Gemini 3 Pro, its market share on the web version has reached a new high, breaking through the 15% mark.

Meanwhile, Grok has also continued to grow with the release of version 4.1.

ChatGPT still leads by a large margin in terms of traffic, but its share is continuously declining.

Gemini has continued its growth momentum, with the number of visits reaching 1.351 billion, a 14.3% increase compared to October. ChatGPT has fallen below the 6 - billion mark, and the website visits have dropped to 5.844 billion.

This is the second month - on - month decline of ChatGPT in 2025 after July.

Although on November 18th (the day of the Cloudflare outage), ChatGPT set a historical record of 233 million visits in a single day, this did not reverse the overall downward trend of traffic.

In addition, the number of visits to the Grok website reached 234.4 million, setting a new historical high, a 14.7% increase compared to October.

This is also the first time that the website has achieved two consecutive months of traffic growth since it started using the current domain name.