Google Physics is waking up! A new Gemini 4 Argon wave test has arrived.

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

Google's Gemini 4 Argon is really here.

Google, once thought to have fallen out of favor, is finally back.

Based on the run score, it plays back and forth with the GPT-6 Astra and Claude Opus 5.5 in text and logic.

In terms of price, it made only one-fifth of its opponents and was several times faster!

What's even more amazing is that it's the first general-purpose model of physical reasoning that makes sense "with images, video, language and even physical intuition at the same time".

Someone shouted: "Gemini 4 Argon is starting to understand physics!"

Ethan Mollick, a professor at the Wharton School of Business, said: "Gemini 4 is very different from other models, and the other models are somewhat similar to Claude/GPT".

What is the effect of Gemini 4 Argon?

Test 1: Variable number of "handrail" cycling rails

In the past, large models have been used to measure the "SVG" for cycling.

But on X, one blogger observed that the difficulty was overwhelming: it allowed Argon to rotate the cylinder wheel to produce a version of a cycling bicycle.

His feeling was that Argon had an amazing intuitive understanding of the world, adopting the principle of primacy.

We also made an observation with Aagon, and surprisingly, Argon instantly realized the geometrical nature of the Zodiac number approaching the closed curve!

At the end, it "calculates" directly out of the steering wheel.

Test 2:3D according to the County Towers

Next, we let it generate a 3D version of the appropriate County Towers.

It looks gorgeous, but it's not shocking enough, and it's not really a proper tower. We just gave it a hint.

The original five-tier structure, now only four layers can be distinguished.

This time, the 3D graphics it generates are good.

We want it to generate a vivid explanation to build a video that has a big feel.

The video that it generates is self-explanatory or in Chinese, but listen carefully and it sounds a bit funny.

Test 3: Deep-sea exploration of the Titanic

Recently, Argon has become so obsessed with all sorts of AI-related videos that he can't get enough of them, that he's created one of the most impressive AI videos I've seen recently.

Let the robot take us to the deep sea and restore the site of the Titanic that year.

Both versions of the video generated only 10s, and could not produce longer videos. If the requirements were not high, the physical quality was available.

In addition, the Chinese accent is still funny.

Testing 4：Big test of scientific reasoning

Gemini 4 Argon has reportedly made many breakthroughs in the field of hard-core science, such as completing quantum computation, solving nuclear fusion problems, diagnosing rare diseases with chromosomal scales, designing five highly active new enzymes, and more.

So let's see how strong the scientific reasoning capabilities of Gemini 4 Argon are.

Tri-dimensional stellar superconducting wire with nuclear fusion magnetic field solution

Since Argon has found a new counter-example of the "Grad's conjecture" of nuclear fusion, we have directly enabled it to manipulate a three-dimensional stellar magnetic field solver using pure code.

This directly reflects Argon's cutting-edge computational physics and three-dimensional differential geometry: it can balance the magnetic fluid mechanics equations and the spatial topology of the superconducting sphere into visual simulations.

Then we replicated the highly acclaimed test of how the Tesla three-phase interchangeable sensing engine works.

This shows that Argon has a very strong multi-physical field (electromagnetic-mechanical) reasoning ability.

In the demos it produced, not only did the three-dimensional mechanical structure align with the axial alignment of the right-handed spiral, but also the three-phase phase difference with the Lorentz force torque causal chain.

Unfortunately, we wanted to test more examples, but as a $20-a-month Pro user, we haven't tested a few cases yet, and our quota is at the top.

What follows are the observations of other netizens.

As soon as the Gemini 4 Argon went live, Case was flying everywhere.

For its 3D generation and spatial visual semantic reasoning, most netizens have praised it.

@Bhavani_00007 Netizens @Bhavani_00007 Argon created a realistic 3D library with just a hint.

His assessment was: This is a performance monster!I never expected any big model to have this much realistic detail.

Indeed, in the case of many netizens, Argon is creative enough in 3D and complex interactive coding.

Argon has been used to combine Nano Banana with three.js to create a futuristic 3D armed helicopter.

The developer made it "a little cool in a.html file", and Argon directly released four physical-level visual effects that looked cool!

It can also directly replicate my World Cup.

Netizens @notjazii spent 19 minutes, consuming 6.3 million tokens and only $3 in Devin, to create the perfect Minecraft clone for Argon.

Some have even made it an impossible museum.

@pankajkumar_dev has tested 3D scene generation, calling it "its 3D speed and creativity is completely another dimension".

One of the biggest advantages of Gemini 4 is that it has a different personality and aesthetic style than Claude and Astra.

What really satisfies users is the extreme price of Argon.

Let's look at a few groups of opposites.

Round 1: Argon vs. Claude Opus 5.5 (Fruit Garden 3D test)

Developers @marcthecreatorr tested it with the same pointers.

Opus 5.5：It takes about 30 minutes and costs about $12.50.

Gemini 4 Argon：It takes 15 minutes and costs about five dollars.

As a result, Argon is not only richer in color, more detailed, and nearly four times faster, at less than half the cost!

Another developer @notjazii came to almost the same conclusion: Opus 15 minutes/8.2 knives vs. Argon 13 minutes/1.2 knives, calling out, "The Google team did something really good this time".

Round 2: Argon vs GPT-6 Astra (performance and cost pressure)

The results of @xKaka1456 are even more exaggerated.

Astra：It takes 25 minutes and costs 11 dollars.

Argon：It takes 3 minutes and costs $0.5.

It's more than eight times the speed and 22 times the cost.

On visual basic tasks (such as target detection, dense scenes, small text/handwritten text/complex form extraction), some netizens declare: "Astra has been kicked out of the temple, Gemini 4 Argon is the best visual model at the moment and is four times cheaper than Astra in High Mode".

Google's official position on Gemini 4 Argon is this:A general model of physical reasoning。

It's designed to open up physical understanding, code generation, and visual semantic reasoning in robotics.

In an ERQA benchmark test with physical reasoning, Argon took a high score of 87% with pure reasoning alone, pressing down on all front-end models.

This means that AI can control physical entities directly in the real world by watching videos.

Case 1: Robot learning to throw sandbags at humans

Google made Argon watch a regular video of humans playing "throwing sandbags".

Argon then mapped the environment himself, calculated the movement of the joints throughout the body, and then independently constructed a simulation code for crazy practice in the virtual world.

Finally, when the data was transmitted to the Apollo humanoid robot, the robot hit seven consecutive shots in real-life testing!

As big V@StockSavvyShay says:

If robots could learn their physical skills from ordinary video, then Google's massive YouTube video library would be the biggest fortress it could have built without a robot in the era of physical AI.

Case 2: From hand-drawn sketches to precise physical throwing

The developer only needs to draw a rough sketch on paper (e.g. a robot throwing a ball over a box on a table).

Argon is able to automatically construct GPU-accelerated MuJoCo physics scenarios, synthesize a perfect 76 degree full-body + five-inch throwing trajectory, verify the physical laws of the aim in a simulation, and ultimately export high-definition rendering video.

In the process, AI fully understands gravity, momentum and collision!

Case 3: A closed-loop visual detects an unusually heavy ball.

Faced with the classic "12 Little Balls" puzzle, Argon was able to write a purely visual closed-loop controller himself.

It controls the two Franka mechanical arms, observes the slope of the plane with only the RGB camera, and with its own planned steps, with just three weights, it accurately picks out the ball of different weight.

From zero-sample machine dog control, to the automatic conversion of human first-person video into robot action, and to the very high-precision video value function evaluation (which determines whether the human operation in the video is effective or incorrect)...

In this way, Argon crosses the physical divide between numbers and reality.

Of course, some hard wounds were also exposed in the Argon test.

Loop One: Drawing on the "real business code"

Although Argon has been criticized for dealing with authentic, highly engineered everyday code, it has also received negative reviews.

Some people say: "This model is useless when it comes to real-world code. Maybe Google has abandoned the competition for a purely programmed model".

Another networker, Argon, made a video introducing himself, which resulted in both versions being extremely bad, with Opus 5.5 doing "One Heaven and One Underground".

Location 2: Anti-human UI and eco-closure

"The model is definitely SOTA and can be higher than the Astra, Opus, but the problem with Google is its product culture and workspace, Harness".

Antigravity's workspace is overwhelming: it's impossible to process logins (which makes it impossible to access X, Reddit or YouTube), it's difficult to browse files generated, and it's only possible to use 40 collaborative apps, all of which are locked down. You can't even log multiple Gmail accounts simultaneously!

Google is in desperate need of a modern UI that matches Anthropic.

Lock 3: Pricing is silent

Some angered that the Gemini 4 Argon is exclusive to Ultra users in Antigravity, and Pro users who pay $20 a month can only use it in regular Gemini chatbots.

However, both OpenAI and Claude offer the best models at a $20 rating.

Over the past few months, the GPT-6 Astra and Opus 5.5 have taken up almost all the attention.

Today's battle, Google is back.

It's not a perfect programmer, but it's actually a "physical-intelligent brain" with physical intuition and top visual spatial capabilities.

And the price of sex is too high.

As Ethan Mollick puts it:

It is unclear who will ultimately win the Battle of the Three Kingdoms.

But what is certain is that we are witnessing the fastest acceleration of physical AGI!