The Gemini 4 Pro has just been released, testing the opus 5.5!

  It's the first time in the history of the world.

Gemini 4 Pro, a new checkpoint has been revealed!

Recently, two brand-new checkpoints -- code-named "barium-b" or Checkpoint 2) -- were discovered by developers who tested and found Gemini 4 Pro to be terribly powerful!

One developer shouted, "After 400 hours of Opus 5.5 and GPT-6, Gemini 4 Pro still makes my hair curl!"

Google's new generation of Big Killers showcases powerful 3D physical simulation, code generation and long-text reasoning capabilities.

One developer shouted, "After 400 hours of Opus 5.5 and GPT-6, Gemini 4 Pro still makes my hair curl!"

In addition, the attitude of the senior management within Google is also evident.

Koray Kavukcuoglu, new head of Google DeepMind, has publicly confirmed for the first time that Gemini 4 has entered early training.

Its training progresses significantly in advance, taking only two months, and early pre-training is basically over.

Get ready, a hard wave of verification is coming!

Overnight, it was as if everyone could use the Gemini 4 Pro.

From the demos unveiled, its performance in 3D generation, physical engine simulation and long-text code building is surprising.

The Falcon SVG test also showed a huge leap forward.

Renowned test blogger @AI_Screening asked it and Opus 5.5 to code to create a physical simulation of a 3D float plane sliding over water.

Previously, the video generated by Opus 5.5 shocked a lot of people, but I didn't think, this time, the performance of the Gemini 4 Pro completely exceeded it!

As you can see, the aircraft produced by Opus 5.5 is moving, but the body is vibrant, as if to break down at any time, and the physical feedback on the water surface is hard.

However, the Gemini 4 Pro is much smoother, faster and smoother, reflects clearly on the water, and is realistic in its rendering of the "hot water" effect.

In terms of fluid mechanical logic and image integrity, Opus 5.5 can be used for ground friction.

The developer @thtbee_ threw out a king bomb.

With no 3D model input, it took Gemini 4 Pro only 16 minutes to create an interactive 3D simulation of the International Space Station running around the Earth using pure code!

Surprisingly, not only did it build its own model of the space station using Three.js, but it also cleverly grabbed the right map of the Earth from the internet, ensuring absolute accuracy of its density and color.

This "non-biological" modeling ability is amazing.

Of course, some people have also spat out some ugly UI interfaces with bugs on the back.

Netizens @HarshithLucky3 gave it a high-difficulty instruction: "Use Three.js to turn a lighthouse into a rocket launch, without any UI elements".

As a result, the Gemini 4 Pro follows the instructions perfectly!

The testers praised its "aesthetics": it adds excellent texture to all the elements in the image, and the overall 3D output quality is good enough to be outspoken.

Developers @LuminaBench have tested its industrial design modeling capabilities to generate a 3D Nintendo Wii remote control.

As a result, it has a very strong grasp of the details of 3D objects and produces "incredibly fast" speeds!

What's not perfect is that it seems a little over-sculpted when dealing with Prompt, and will do its best to enrich the details, even if you don't ask for it.

However, its design aesthetics are not very stable in different parts, some places excellent and some slightly inconsistent.

According to the tests, the detail density of Checkpoint 2 is excellent, with much higher accuracy, and is the fastest front-end model currently available.

Blogger @TimJayas has released his "Patriot" machine bee birdprompt, which tests the model's understanding of complex mechanical structures and dynamic behavior.

The conclusion was: "The performance of the Gemini 4 Pro is amazing! Honestly, the quality is similar to that of the Fable flagship, but far better than the GPT-6 Sol!"

Of course, it is sometimes pressurized by the GPT-6 Astra.

In addition, its engineering capabilities are incredible.

In just 14 minutes, a developer created an incredibly creative product to showcase a complete website, front-end and back-end.

What's even more surprising is that a lot of people give feedback, and you just need to type in a Prompt, and it gives you a complete interactive web game.

While new checkpoints have only recently been revealed, the parameters of the Gemini 4 Pro have already been leaked.

It has been revealed that the Gemini 4 Pro has an upper context window of 10 million tokens, and an output limit of 256k, if true, would be a complete farewell to the nightmare of "writing code halfway through", a phrase that generates the entire code set of large software.

It has a permanent cross-session memory, connectivity to the internet without an API, and real AI Agent capabilities.

It even has the ability to control a robot directly as a brain that controls a robot in the real physical world!

A previous benchmark test showed that the Gemini 4 Pro outperformed the GPT-6 Astra and Claude Fable 5.1 in programming, intelligence and reasoning.

It took 88% on the AI programming test (DeepSWE v1.1) and 95.3% on the coding capability test (Terminal-bench 2.1)!

Some people shouted: "I smell the RSI".

Why did the Gemini 4 Pro crash suddenly? Because Google's bottom line was hit.

After some time behind GPT-6 and Opus 5.5, the giant began to react.

It was only on Wednesday that Google DeepMind's new boss, Koray Kavukcuoglu, made his first public disclosure of "crossroads", revealing a major news story.

According to The Information, Koray revealed that Gemini 4 has been in the early stages of training for about two months after launching "the most ambitious pre-training ever" in July.

At the moment, Gemini 4 is fully entering an important "post-training" phase.

"Our intention is to release an early post-training version as soon as possible", Koray said with enthusiasm, "because we're very excited to see the results!"

This is why the test node "half-finished but already very strong" will suddenly appear on the arena.

Moreover, after finding that the Gemini 3.5 Pro was not performing well enough to put pressure on its opponents, Google cut its wrist and turned all its computing power to Gemini 4!

When asked sharply whether he was concerned that Google was lagging behind its peers, Koray said sharply: "I have 100% confidence in my team and we are destined to always be at the forefront of technology".

He revealed that Gemini 4 is currently being used by internal engineers for Antigravity AI, a response to earlier rumours that "Google employees write code with Claude".

And, he stressed, they no longer cling to the concept of AGI, "The real question is, do we have the ability to build an intelligent body that we can fully trust?"

He also revealed that the development of Gemini 4 helped Google hardware engineers design "the next two to three generations of TPUs".

Finally, Koray tossed out a timetable: he hoped the official release date for Gemini 4 would be "much earlier than the end of the year".

There are also reports that the Gemini 4 Pro 11 could be released.

Even if the performance explodes, Google is preparing for a price war. According to the leak, the pricing of the Gemini 4 Pro is highly aggressive:

Input: $2.25 / 1 million Tokens

Output: $11.25 / 1 million Tokens

Compared to the flagship model of dozens of American knives traded by friends, this is simply a "bone-breaking price".

It's no wonder people are shouting: "The price of sex is gone, the 4 Pro is too much to ask for, and it's too late to become a member!"

And Google is going to send the AI chip to heaven!

Next week, with SpaceX's Falcon 9, Google will launch a satellite equipped with four TPU chips into space!

In the future, there may be a huge network of AI data centers in the sky.

It seems that the Gemini 3.8-flash in the previous LMSYS is just the tip of the iceberg of "extremely early post-training".

But as the testers put it: "The second checkpoint is a huge leap from the first checkpoint, and Google is really on a roll this time around".

What is the relationship between OpenAI and Anthropic?

Will the addition of Gemini 4 Pro change the outcome of the ASI Final?