Gemini 4 Argon in Practice: 3D Scene, Aquarium, Game, and Design

In this review, Gemini 4 Argon is a model the author tested by generating interactive projects through OpenRouter. The test used High mode, which the author describes as the highest reasoning level. The projects included an 3D orc model, an aquarium scene, a Super Mario-style game, and a website design.

The results show not the model’s universal capabilities, but one author’s specific experience with given prompts and a chosen access method. The author assessed appearance, animation, logic, and generation stability separately. A successful feature in one project does not imply the same quality in all the others.

Gemini 4 Argon and Testing Conditions

The author ran Gemini 4 Argon through OpenRouter, selecting High mode. They note that providers sometimes experienced outages during testing. According to the author, Google Vertex and Google AI Studio worked reasonably well during this period, and they did not use third-party providers for comparison.

This distinction matters when interpreting the results: an observed delay or interruption could have been related to the access route and provider performance, not just the model’s behavior. The test cannot establish the exact cause of each failure. These results can only be compared with tests of other models with a caveat: the launch conditions and scenarios may have differed.

3D Orc Model: 17 Minutes and 38 Cents

According to the author’s measurements, Gemini 4 Argon took 17 minutes and 38 cents to create the 3D orc model. The author did not consider the model particularly high-quality: its detail and overall execution did not convince them. However, one specific element was done correctly: the orc held the axe with its blade facing forward.

The author rated the animation more highly than the model itself. In their view, the movements worked normally and conveyed noticeable body language, although the character’s visual polish fell short of some alternatives they compared it with. This observation relates to one scene and does not measure the model’s overall 3D generation quality.

For a prototype where basic movement and the correct orientation of an object matter, this result may be useful as a rough draft. If a project requires an expressive model, the author’s assessment suggests that additional artistic refinement would be needed. The cost and time apply only to this run.

Aquarium: Working Animation and Noticeable Flaws

The aquarium took about an hour and a half to create and cost the author 4 dollars and 75 cents. During generation, the model stopped about six times, and the author had to send the command “Continue.” They attributed the repeated stops to unstable provider performance through OpenRouter, but that was their explanation, not a confirmed diagnosis of the cause.

The scene featured a crab, two snails, and fish. According to the author, the crab walked and moved its claws, while the snails and fish looked fairly good. The bubbles and aerator worked, the image distorted under the effect of water, and the fish animation seemed fine to the author.

The setting had noticeable flaws: the castle looked unnatural, and the bottom surface resembled tile, although sand was expected. The author called the aquarium functional but unremarkable and considered the cost high. Their suggestion that the result might have been better in Google’s native environment remained a supposition: the test did not compare the same project across different environments.

This scenario illustrates the difference between having interactive elements and having a cohesive visual scene. The fish and crab moved, but the setting details did not always match the expected appearance. That may be enough for an initial demonstration; based on this example, the author would have needed to review and edit the finished scene.

Super Mario-Style Game: Stronger Logic Than Presentation

Gemini 4 Argon created the game in 26 minutes, and the author noticed no interruptions during this run. However, they could not open the game simply by running the file: the author started a local server and was then able to test the project. This was a condition of this specific test, not a claim that every project from the model requires a server.

The game included instructions, coins, a mushroom, character growth, enemies, pipes, a secret area, an additional level, and a checkpoint. The author tested several interactions: the character moved through pipes, reacted to mushrooms, shot fireballs after getting a flower, and the turtle recovered from its hibernation state.

The author rated the game logic as good, but the enemies’ behavior was inconsistent: they sometimes jumped into unsuitable places on their own, and some pipes did not work. The appearance also needed refinement. Mario looked blocky, and the character’s face was visible only from one side.

For vibe coding, this example is interesting because it combines working mechanics with visual limitations. The project already made it possible to play through the scenario and test several features, but that did not make it a finished game. Further use would require testing the levels, checking the pipes, and fixing enemy behavior.

Website Design: Expressive but Unclear

The design project cost 1 dollar and 24 cent. The author expected it to take about 50 seconds, but it continued longer because of repeated interruptions. They again attributed the delays to provider performance on OpenRouter; the results of the described test cannot confirm this cause.

The author called the resulting interface beautiful but difficult to understand. It included color settings, various controls, and additional menus, but the purpose of many elements remained unclear. Visual flair alone did not help explain what the website offered or what action a visitor was supposed to take.

The studio website looked unusual, but the author could not determine its purpose or main message. After loading it, they also found no responsive version: according to their observation, the website appeared on an iPhone in its original layout. They drew the same conclusion about other formats based on their test. The interface was able to calculate some cost, but the author did not specify exactly what calculation it performed.

For a designer or developer, an eye-catching first screen is not enough. It is important to check whether visitors understand the page’s purpose, whether the interface is readable on a mobile device, and whether the controls work as expected. In the project described, the author saw visual flair but found no clear structure or responsive display.

How Much Did OpenRouter Affect the Results?

The author linked the repeated stops during the aquarium and website design generation to unstable provider performance through OpenRouter. During testing, they said, Google Vertex and Google AI Studio worked reasonably well. However, this does not mean that OpenRouter always reduces Gemini 4 Argon’s quality, or that switching to Google’s environment would necessarily improve the results.

It is difficult to separate the possible influence of the provider from that of the model itself: the test does not include a controlled comparison of the same prompt across several environments with identical settings. The author suggested that the aquarium might have turned out better in Google’s native environment. This is their hypothesis, not an established experimental finding.

Comparison with Claude Opus and GPT-6 Astra

In the author’s subjective assessment, Claude Opus and GPT-6 Astra handled the same projects and prompts noticeably better. This comparison reflects their impressions of the runs they conducted, not a standardized benchmark: the account provides no single numerical quality scale, repeated test series, or detailed parameters for each competitor.

The author acknowledged that the difference could have depended on system load or on how well the chosen projects suited the particular models. They planned to test other tasks and conduct additional comparisons. Until those tests are done, it is more accurate to discuss the results of these specific runs rather than claim that one model is categorically superior to another.

Conclusion: Useful for Prototypes, but Check the Results Manually

In the tests, Gemini 4 Argon demonstrated several useful features: it animated an orc, assembled an aquarium with moving creatures, and created a game with several gameplay mechanics. What impressed the author most was not the visual result but the game logic. At the same time, they found the 3D orc model weak, the aquarium had unnatural details, and the website looked impressive but did not explain its purpose.

What does practical generation cost? In the runs described, the orc cost 38 cents, the aquarium cost 4 dollars and 75 cents, and the design project cost 1 dollar and 24 cent. These amounts apply to specific tests and do not establish a rate or typical cost for future projects. The time also varied: 17 minutes for the orc, about an hour and a half for the aquarium, and 26 minutes for the game; the design took longer than the 50 seconds the author expected because of interruptions.

The author came away with mixed impressions and considered the visual quality below their expectations. For a developer interested in quick prototyping, testing mechanics, and manually fixing the result, the test offers reason to try the model on their own task. If a project requires a finished mobile website, polished 3D graphics, or stable, long-running generation, the examples described show that both output quality and provider performance should be checked separately.