Skip to content
Original
Xiaolin coding·· 1 day agoSelectedAI score74

Gemini 4 Argon launches: code-generation results and pricing compared

AI overview

Gemini 4 Argon is now generally available. It scores 53 on Artificial Analysis, level with Claude Fable 5.1 and GPT-6 Astra at Max, and ranks fifth overall on Agent Arena.

Why it matters

A roundup of hands-on examples from multiple developers and pricing comparisons shows where Gemini 4 Argon stands on code generation and cost.

Full text · AI translation

Original by Xiaolin coding 2026-10-14 17:06 Guangdong

Illustration accompanying “Original by Xiaolin coding 2026-10-14 17:06 Guangdong”

Hi everyone, I'm Xiaolin.

After falling behind for half a year, Gemini is finally done being a cute “Hakimi.” This time it has become a dragon: Gemini 4 Argon has officially launched!

Illustration accompanying “After falling behind for half a year, Gemini is finally done being a cute “Hakimi.” This time it has become a dragon: Gemini 4 Argon has officially launched!”

It has immediately returned to the top tier, trading blows with GPT-6 Astra and Claude Opus 5.5.

Artificial Analysis's model page gives Gemini 4 Argon (High) a score of 53, level with Claude Fable 5.1 and GPT-6 Astra at the Max setting.

Illustration accompanying “Artificial Analysis's model page gives Gemini 4 Argon (High) a score of 53, level with Claude Fable 5.1 and GPT-6 Astra at the Max setting.”

On the overall Agent Arena leaderboard, Gemini 4 Argon (High) ranks fifth, behind GPT-6 Astra (Max) and ahead of Claude Sonnet 5.5 (Max).

Illustration accompanying “On the overall Agent Arena leaderboard, Gemini 4 Argon (High) ranks fifth, behind GPT-6 Astra (Max) and ahead of Claude Sonnet 5.5 (Max).”

Gemini 4 Argon was actually announced on 9/30. Readers were already urging me to test it, but at the time only some enterprise users had access; it was not generally available.

Illustration accompanying “Gemini 4 Argon was actually announced on 9/30. Readers were already urging me to test it, but at the time only some enterprise users had access; it was not g…”

As soon as the full launch happened, I opened Google Antigravity to try it. Gemini 4 Argon was indeed listed, but then I discovered it was available only to Ultra subscribers. Awkward.

Illustration accompanying “As soon as the full launch happened, I opened Google Antigravity to try it. Gemini 4 Argon was indeed listed, but then I discovered it was available only to…”

As luck would have it, I subscribe to Pro. Pro users can only use Gemini 4 Argon in the Gemini chat app, not in Google Antigravity.

So Pro users can only write code through the chatbot? Does that make sense? It feels a bit like Pro users are being treated as second-class subscribers…

Still, I have already seen plenty of experts testing Gemini 4 Argon on X. Let's take a look at their results.

────Gemini 4 Argon in practice────

A plush pumpkin with 6 × ten thousand hairs moving in the breeze

How convincing can a plush toy look when it is made purely with code, without a modeling tool such as Blender?

Developer @thebuggeddev used Gemini 4 Argon to make a 3D pumpkin toy whose fur responds to wind in real time. According to the developer, the model was generated with code, without Blender, and has 6 × ten thousand hairs.

Just look at that soft texture. Doesn't it make you want to reach through the screen and give it a good pet?

Illustration for the section “A plush pumpkin with 6 × ten thousand hairs moving in the breeze”

It is animated, too: wherever the wind blows, the fur moves with it.

Illustration for the section “A plush pumpkin with 6 × ten thousand hairs moving in the breeze”

How is this even possible? Gemini 4 Argon made this simply by writing code…How is this even possible? Gemini 4 Argon made this simply by writing code…

Art through code: 534 works, some even animated

Even more astonishing were the 534 artworks Surya (@kickingkeys) shared on X, all drawn with code written by Gemini 4 Argon. To be clear: they were drawn with code, not generated by an image model.

They range from animals and everyday objects to abstract patterns, with moving windshield wipers and falling snowflakes. Some pieces are animated as well.

Illustration for the section “Art through code: 534 works, some even animated”

At first I did not believe images like these could be drawn with code. Then I saw that the creator had published the source for every piece. I opened it, confirmed they really were code-generated, and had my expectations reset once again.

Illustration for the section “Art through code: 534 works, some even animated”

It is almost absurd that code-generated images can have such a strong sense of texture. Perhaps Gemini 4 Argon could even match Opus 5.5 at producing video effects.

A music synthesizer with a keyboard, sequencer, and live sound controls

This time, the task was to bring an electronic synthesizer into the browser.

@maxescu used Gemini 4 Argon to build a music synthesizer combining Jupiter and Moog styles. According to the creator, it took less than half an hour, and the entire application fits in a single HTML file.

Illustration for the section “A music synthesizer with a keyboard, sequencer, and live sound controls”

In the demo, the keyboard lights up with the melody, notes and drumbeats play together, and you can turn knobs to change the sound during playback. Definitely watch this one with sound on.

A Minecraft clone with water reflections and travel to the Nether

Now for games. @notjazii asked Gemini 4 Argon to make a Minecraft clone. According to the creator's records, this version took 19 minutes and cost about 3 US dollars.

Illustration for the section “A Minecraft clone with water reflections and travel to the Nether”

There are wooden cabins, grass, and pink tree canopies, with the surroundings reflected in the water. The hotbar, health, and hunger indicators are all in place at the bottom.

In the demo, the player can fly and switch the item they are holding. Going through a purple portal changes the scene from grassy cabins to the red Nether.

This is a remarkably complete implementation. I have tried building Minecraft with other models before, but none of those versions were this complete.

SVG animation comparison: day/night switching and following eyes

How different can the visual styles be when two models get the same animation task?

@Adidotdev gave Gemini 4 Argon and GPT-6 Astra the same set of SVG animation prompts. According to the creator, Argon took about 5 minutes per task. Argon is on the left of the video and Astra on the right.

First, the day/night toggle. Argon's sun and moon have distinct light and shadow, with clouds, stars, and background changing together during the transition. Astra's version is simpler, focusing mainly on the switch itself.

Illustration for the section “SVG animation comparison: day/night switching and following eyes”

Next, the chameleon. Argon's body is rounder, with highlights and shadows on the eyes that create a stronger sense of depth. Astra's outline is slimmer, with clearly defined stripes and spots, giving it more of a flat illustration style.

Illustration for the section “SVG animation comparison: day/night switching and following eyes”

Which result do you prefer, Argon's or Astra's?

A dragon and an ancient tower: bodies, buildings, and details from rotating viewpoints

Modeling used to be where Astra really impressed me. Now it feels thoroughly outclassed by Argon.

@marcthecreatorr shared this voxel-scene comparison, saying both versions used the same prompt. Argon is on the left and GPT-6 Astra on the right.

Illustration for the section “A dragon and an ancient tower: bodies, buildings, and details from rotating viewpoints”

In the rotating view shown earlier, Argon's dragon has a more contoured body, with clearly visible back spikes and spread wings. You can also see the tower's window grids, railings, and upturned eaves. Astra's design is considerably simpler.

A space train: a floating station, moving light and shadow, and time comparisons

Can Argon match Opus 5.5 at making videos?

@EnvolDev used the same prompt to have Argon and Opus 5.5 each build a space train. Argon is above and Opus below.

Illustration for the section “A space train: a floating station, moving light and shadow, and time comparisons”

Argon's vaulted station floats in midair. A train runs along the tracks trailing white smoke, with warm light shining from its windows. The camera tracks toward the locomotive as lightning strikes nearby.

According to the creator's records, Argon took 4 minutes 20 seconds and cost 3.40 US dollars. Opus 5.5 took 1 hour 53 minutes and cost 23.46 US dollars.

In terms of the final result, Opus 5.5 is admittedly better, with strong locomotive detail and camera movement. But when time and cost are considered, Argon decisively beats Opus 5.5 on speed and value.

A cycling pelican takes its ride onto a fighter jet

You have surely seen a pelican riding a bicycle. But have you seen one riding a fighter jet?

@VulKan42069 on X shared an even wilder comparison: a pelican doing a wheelie on a bicycle, perched above a fighter jet in flight. Argon is on the left and Opus 5.5 on the right.

Illustration for the section “A cycling pelican takes its ride onto a fighter jet”

Look closely at Argon's version on the left. The cockpit, wings, and fuselage have richer detail, and an orange-yellow particle exhaust trails behind. As the camera moves around to the side, you can clearly see the pelican's long beak, red scarf, and raised bicycle wheel. Opus 5.5's aircraft and clouds on the right are simpler.

What an imagination. Is the pelican going to space next? Hilarious.

A rolling ball sets off a mechanical chain reaction

Argon's physical reasoning is also very strong.

On X, Matthew Berman shared a 3D chain-reaction mechanism made with Gemini 4 Argon, also known as a Rube Goldberg machine. Once the ball moves, the mechanisms activate one after another like a relay.

Illustration for the section “A rolling ball sets off a mechanical chain reaction”

In a red canyon, wooden frames support tracks at staggered heights. In the demo, a ball rolls down a ramp, a row of dominoes falls in sequence, and a hinged panel tips with them.

The camera follows the mechanism downward while the stage indicator updates alongside it. A panel on the right lets you adjust gravity, friction, and bounce, like a small physics lab.

────Final thoughts────

Finally, the pricing everyone cares about: Argon blows the competition away here, too.

The price is just 2 US dollars per million input tokens and 10 US dollars per million output tokens, with a discount of up to 95% for cached input. Argon is 50% cheaper than Opus 5.5 and 80% cheaper than GPT-6 Astra, while also being faster.

These are introductory prices, however. The offer lasts until 11/30, after which standard pricing returns at 4 US dollars / 20 US dollars.

Looking at price and capability together makes the AA chart below especially useful. Higher means a better score, while further left means a lower average cost to complete the test tasks. Argon's average cost per task is marked as 1.99 US dollars. The competitive cutoff is moving astonishingly fast.

Illustration for the section “────Final thoughts────”

After being out of the running for half a year, Gemini has returned to the top tier. I hope Chinese models can deliver a surprise like this too.

Source: Xiaolin coding · mp.weixin.qq.com