Gemini 4 Argon is the cheapest frontier model in this test and still lost every build. In four tasks run in October 2026 it stayed fast and inexpensive while Claude 55 produced better visuals at up to ten times the cost, so price alone does not settle the choice.
What Gemini 4 Argon is, and what the review actually measured
Gemini 4 Argon is Google's frontier model released to the general public in October 2026, and the review compared it with Claude 55, Fable 5.1 and GPT-6 Astra on four long-running builds. Each model ran at high effort on the same prompts: a browser 3D coffee bottle, a launch video, a Rocket League-style car soccer game and a pixel-level website clone.
The reviewer is Pat Simmons, publishing a 45-minute hands-on review on 14 October 2026. The method matters more than any single score: every model built in headless Blender through Python scripts, had access to image generation, and ran one shot per task. No task was retried, and no knowledge-work task was tested despite Google's claims there.
Four tasks is a small sample, and the reviewer says so. He picked heavy 3D, video and game work because it is hard to fake, not because it matches most people's day. Treat the results as a signal about visual and physical realism, not a full product evaluation.
Google's benchmark claims versus one creator's builds
Google's published benchmarks put Gemini 4 Argon ahead in several knowledge-work and agentic categories, yet the review's builds did not reproduce that lead on visual tasks. The reviewer read two Google posts, one of demos and one of benchmarks, and noted the demos leaned heavily on 3D work: temple complexes in Blender, a desktop car race, a 3D chess set and a ray tracer with an audio mixer.
Two benchmark results stood out in the review. On Harvey's Legal Bench, a legal-reasoning evaluation, Gemini 4 Argon reportedly scored 19.6% against 5% to 6% for rival models, and on an automation benchmark it reached 51% against the 30s and 40s. Those are Google-reported figures, and the reviewer treats them with caution.
The weakest result for Argon was Terminal Bench, a benchmark of day-to-day terminal coding tasks. The reviewer, who uses Terminal Bench as a proxy for everyday coding, said Argon was not competitive with Claude 55 there. Anyone deciding on coding work should weigh that result more heavily than the 3D demo renders.
The reviewer also compared a prior release from the same model line, which he recalled scoring around 79 on the SWE-bench family of coding benchmarks, to show that a high benchmark score can coexist with a weak coding agent. That comparison is his recollection of a different review, not a re-run.
Cost per build: where Argon is genuinely cheapest
Gemini 4 Argon was the cheapest model in every task, and by a wide margin in three of the four. The figures below come from the reviewer's own cost tables and are API-equivalent estimates using promotional pricing of $2 per million input tokens and $10 per million output tokens, a promotion running until 30 November 2026.
3D product build: price did not buy realism
The 3D coffee-bottle build exposed the clearest gap between Argon and the field. Every model received the same strict prompt: a photoreal interactive bottle in the browser, glass with correct refraction and thickness, a printed-paper label, and a click-to-pour interaction. Claude 55 won on realism; Argon placed fourth.
Argon's bottle and glass looked close to photoreal at first glance, but the pour animation looked unnatural and the render briefly went black mid-rotation. The reviewer ranked it below an entry that used a stylized illustration, because that entry still read as a considered design rather than a failed photoreal attempt.
Argon cost $6.85 and finished in 36 minutes. GPT-6 Astra cost $8.99 in 34 minutes, Fable 5.1 cost $17.96 in 57 minutes, and Claude 55 cost $29 while taking 114 minutes. The win was bought with time and money, and the reviewer judged that the better output justified it.
Launch video and game build: polish and physics
The launch video and the Rocket League-style game tell the same story at different sizes. In the launch video, Claude 55 built a connected sequence with a cinematic reveal; Argon delivered a fast, aggressive cut with dense on-screen text and lost product detail. In the car soccer game, Opus produced a match that came close to Rocket League's feel, and Argon's version felt less controlled.
The launch video cost table shows how different the working styles were. Argon spent $10 across 17 million input tokens in 80 minutes. GPT-6 Astra spent $28 across 19 million input tokens in 81 minutes. Fable 5.1 burned 62 million input tokens for $242 over 302 minutes. Claude 55 used 109 million input tokens and 313 thousand output tokens for $109 over 318 minutes.
In the game build, Argon cost $10 in 54 minutes, GPT-6 Astra $14 in 56 minutes, Fable 5.1 $32 in 77 minutes and Claude 55 $46 in 155 minutes with 143 million input tokens. The reviewer said he doubted four times the turns would bring Argon's game up to Opus's level, which is his judgment rather than a measured result.
Website clone: the one task with a clear craft gap
The website clone produced the widest quality gap, and it was about craft rather than raw rendering. All four models rebuilt a joke e-commerce site built around a 3D coaster. Claude 55 reproduced the opening animation and coaster shading almost exactly; GPT-6 Astra produced a convincing hand and packaging but skipped a coffee-cup animation; Argon produced a flat, glitchy hand and 3D that looked closer to an older model generation.
Cost and time diverged sharply here. Argon spent $11 over 77 minutes. Claude 55 spent $67 over 129 minutes. GPT-6 Astra spent $357 over 499 minutes, nearly eight hours of continuous work that the reviewer linked to its detailed coaster and hand rendering. Fable 5.1 spent $98 over 112 minutes and missed the mark.
This is the section to read closely if you care about 3D in the browser. Argon's output looked like an illustration where the prompt asked for photoreal, and the reviewer called the hand the creepiest part of the run.
What the results mean for picking a model in October 2026
On this evidence, Gemini 4 Argon makes sense as a low-cost exploratory model, not as the default for demanding visual work. Its speed and price are real advantages: it finished every task in 36 to 80 minutes at $6.85 to $11, and it never ran long enough to distort a schedule. The realism gap is the cost of that choice.
The price advantage has a deadline. The reviewer notes that the promotional rate runs to 30 November 2026, after which Argon's price moves toward Opus territory, and he argues that at parity the performance gap is hard to justify. Check Google's current pricing page before budgeting, because promotional terms change.
A second caveat is access through a subscription. The reviewer says he was on a $99 monthly plan and hit daily usage limits after roughly one build, which he describes as faster than Codex or Claude Code. That is one user's experience, not a published quota.
For knowledge work in Google Workspace, Google's benchmarks remain the strongest argument for Argon, but the reviewer points out that Claude ChatGPT now reach Docs, Slides and Sheets through connectors. Verify current connector support in each provider's documentation before relying on that claim.
FAQ
- Is Gemini 4 Argon better than Claude 55? In one hands-on review published in October 2026, Claude 55 won all four visual builds on realism and polish, while Gemini 4 Argon was cheaper and faster. Google's benchmarks show Argon ahead on legal and automation tasks. The right choice depends on whether you need visual quality or low cost per run.
- How much does Gemini 4 Argon cost per build? The reviewer measured $6.85 to $11 per build under promotional API pricing of $2 per million input tokens and $10 per million output tokens, valid until 30 November 2026. Comparable builds on Claude 55 cost $29 to $67. Check Google's current pricing page, since promotional terms expire.
- What did Google claim about Gemini 4 Argon's benchmarks? Google reported a large lead on Harvey's Legal Bench, with 19.6% against 5% to 6% for rivals, and 51% on an automation benchmark. The reviewer treats these as vendor-reported results and notes that Argon lagged on Terminal Bench, a coding-task benchmark he trusts for day-to-day work.
- Should I use Gemini 4 Argon for coding? The review did not run software coding tasks, but its author reads Terminal Bench as a proxy for everyday coding and found Argon uncompetitive there. Test it on your own repository before switching.
- What are the biggest caveats in this review? The sample is four one-shot builds chosen for heavy 3D, video and game work, not a broad workload. The reviewer says so himself and calls the results indicative rather than definitive. Cost figures are API-equivalent estimates, and the price promotion ends on 30 November 2026.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits