Hands-On: Gemini 4 Argon vs Claude Opus 5.5 and GPT 6.1 — Similar Pricing, But Falls Behind Across the Board

Google has finally released Gemini 4 Argon, roughly eight months after its last model update. The model is explicitly positioned as a flagship to compete with ChatGPT and Claude, and judging by the naming and parameter scale, it genuinely has substance. But after hands-on testing, my conclusion is: the model itself is good, but at its price point, it's hard to convince users to abandon cheaper and better-performing competitors.

I set Gemini 4 Argon to high thinking effort via Google Antigravity (Google's official Gemini development environment), while running Claude Opus 5.5, GPT 6.1 Sol, and open-source representative DeepSeek^Ad V4.1 Flash through the exact same prompts, conducting three rounds of head-to-head comparisons spanning a pixel art generator, a 3D flight simulator, and a music visualization tool.

Pixel Art Generator: Left Behind by ChatGPT and Claude

The first task was converting uploaded images into genuine pixel art, rather than simply downscaling with a pixelation effect. Claude Opus 5.5 generated multiple preset styles (8-bit, 16-bit, retro RPG, modern indie, arcade, cozy pastel, etc.), with 16-bit and modern indie styles being the most polished. GPT 6.1 went further, using the ChatGPT Sites feature to deploy the generator as an accessible online site.

Gemini 4 Argon clearly lagged. Both the 8-bit and 16-bit presets were underwhelming, and the modern indie style was even worse. DeepSeek's results were similarly lackluster, with preset styles overall on the rough side. The winner of this round was clear: GPT 6.1 had the most refined visuals, Claude Opus followed closely, and Gemini and DeepSeek were noticeably behind.

One detail worth noting: no image model was used here—the pixelation effect was simulated purely through front-end code, so the differences in results reflect each model's understanding of visual style and code generation capability.

3D Flight Simulator: Playable, But Controls Fall Short of Claude

The second task was more challenging: building a browser-playable 3D flight simulator using Three.js technology, with the same aircraft image as a reference.

Gemini 4 Argon's version was quite playable—the aircraft shape matched the uploaded image, camera angles could be switched, and increasing engine power produced visible engine exhaust flames—a nice touch that earned extra points.

DeepSeek's version had obvious bugs: trees covered the runway, the plane veered left for no reason, controls were difficult, and it was essentially unplayable.

GPT 6.1 Sol had the best sense of visual distance, with a clean map free of glitches, but the controls were heavy and sluggish, and the W key's pitch direction was inverted compared to conventional flight games, requiring adaptation.

Claude Opus 5.5 offered graphics quality tiers (low/medium/high) and time-of-day scene selection, with controls closest to a real flight game—having weight but controllable and predictable, plus the cleanest UI. The winner of this round was still Claude Opus 5.5.

Three.js is an open-source JavaScript 3D graphics library based on WebGL that runs in the browser, rendering three-dimensional scenes without plugins. It wraps WebGL's low-level interfaces, providing high-level abstractions such as scene graphs, cameras, lighting, materials, and geometries, and supports features like shadows, post-processing, and animation loops. In this test, models needed to use Three.js to build aircraft models, runway terrain, camera controls, and physics-style flight operations, so the quality of generated code directly determined the realism of the scene, interaction responsiveness, and whether bugs like clipping or coordinate errors existed. Three.js is also one of the most commonly used tech stacks for evaluating a model's front-end 3D coding ability.

Music Visualization: Chaos vs. Restraint

The third task was uploading audio and generating music visualization effects. Gemini 4 Argon's version was sensitive to beat responses, with visual elements bouncing all over, but the visuals lacked structure—even less pleasing than early 2000s music player visualizations.

GPT 6.1 went to the other extreme: its response to downbeats was too restrained, with almost no visible changes, but the overall visuals were clean, offering color and mode settings for a more comfortable viewing experience. DeepSeek provided a fractal-bloom-like effect with sensitivity somewhere in between, but the visuals weren't particularly attractive.

Claude Opus 5.5's design was the most complex, with sufficient sensitivity to the music and the most refined visual style. Overall, Claude Opus won this round again, while Gemini and GPT were roughly tied—but considering price, Gemini's competitiveness was further undermined.

Pricing and Speed: This Is the Real Problem

Gemini 4 Argon currently enjoys Google's discount price, roughly half of the original, but starting this December, the price will rise to a level comparable to Claude Opus 5.5. According to data from independent benchmark platform Artificial Analysis, Gemini 4 Argon's high-effort mode is roughly in the same tier as Claude Fable 5.1 and GPT 6 Astra's best-effort mode, but lags behind Claude Sonnet 5.5's best-effort mode, Meta AI's Amuse Wonder, and Claude Opus 5.5's best-effort mode.

More critically, Gemini 4 Argon consistently exhibited an "overthinking" tendency during demos: the model itself isn't slow, but it automatically iterates repeatedly and self-tests, meaning a single task could take anywhere from one to several hours to complete. GPT 6.1 Sol is not only cheaper and faster but also performs better in most tests. DeepSeek V4.1 Flash is the value king, priced far below the $1 level, with ample token volume and extremely high efficiency.

Additionally, Gemini 4 Argon's usage quota is calculated separately from other Gemini models, similar to how Anthropic separates Claude Fable from Haiku/Sonnet/Opus in its accounting. The free tier and Google AI Plus subscription cannot use this model—at minimum, a Google AI Pro tier or above is required.

Artificial Analysis is an independent third-party AI model evaluation and analysis organization that conducts standardized testing of various models through a unified API, aggregating metrics such as throughput speed, latency, unit cost, and a composite intelligence index. It does not conduct vendor-sponsored targeted evaluations, so it is widely cited when comparing the actual performance and cost-effectiveness of models from different companies. Because models vary greatly in default parameters, thinking budgets, and pricing strategies, third-party unified benchmarks are often more informative than vendor-reported data—the rankings mentioned in this article, such as "same tier" and "best-effort mode," are based on this kind of standardized testing methodology.

The so-called "overthinking" refers to reasoning-capable models expanding multiple rounds of internal reasoning, self-checking, and iterative correction before giving a final answer. This mechanism can improve accuracy on complex tasks like math and code, but it also significantly lengthens end-to-end time and consumes more tokens. Different models have different strategies for controlling thinking budgets: some converge early when confidence is sufficient, while others tend to verify repeatedly. Evaluations often use tiers like "high effort/best effort" to regulate this behavior, and whether a balance can be struck between quality and efficiency has become an important dimension for measuring the practicality of reasoning models.

Conclusion: Worth Watching, But Not Worth Paying For Right Now

Gemini 4 Argon is a significant upgrade over Gemini 3.1 Pro released eight months ago, and Google has finally returned to the competitive track after a long silence—which is good for the entire industry. But at current pricing and actual performance, Claude Opus 5.5 is superior in most tasks, GPT 6.1 Sol delivers comparable or better results at a lower price, and DeepSeek V4.1 Flash is unmatched in the open-source and ultra-low-price segment.

If you're already subscribed to Google AI Pro or a higher tier, there's no harm in trying Gemini 4 Argon. But if you're choosing among paid models, it's hard for it to be the first choice at this stage.

One-line Summary

Gemini 4 Argon is capable, but its pricing can't compete with cheaper and stronger rivals.

Paragraph Summary

The article conducts three rounds of identical prompt tests, comparing Gemini 4 Argon, Claude Opus 5.5, GPT 6.1 Sol, and DeepSeek V4.1 Flash head-to-head. Across pixel art generation, 3D flight simulator, and music visualization tasks, Gemini 4 Argon lagged behind Claude Opus 5.5 in all three, trading wins with GPT 6.1 but generally falling short. More critically, pricing: Gemini currently enjoys roughly half-price discounts, but starting December it will rise to a level comparable to Claude Opus 5.5, while GPT 6.1 Sol is cheaper and faster, and DeepSeek offers extremely high value at far below the $1 level. Combined with Gemini's overthinking tendency, separately calculated usage quotas, and the requirement for Google AI Pro or above, the author's conclusion is: the model is worth watching, but not worth paying for as a first choice at this stage.