Gemini 4 Argon is the faster, cheaper-looking choice on paper and still lost every judged build to Claude 5.5 in an October 2026 test. The twist is that token totals favored Opus in most projects, and a five-hour restart inflated Gemini 4 Argon's input token counts, so raw cost per project is harder to read than it first appears.
Gemini 4 Argon vs Claude 5.5: which model won the build test?
Gemini 4 Argon lost all six head-to-head builds to Claude 5.5 in a test published on 14 October 2026 by the channel The AI Automators, scoring between 4 and 6 out of 10 against Opus scores of 6 to 8. The builds were judged separately against written specs, so the result measures how much of each brief each model finished, not raw code quality. Gemini 4 Argon is a Google model and Claude 5.5 is an Anthropic model; neither had an edge in effort settings, because both ran at high reasoning effort inside Google Antigravity, the agentic coding environment used for the tests.
The setup matters as much as the score. Each model received the same prompt, a per-project time limit, and one attempt. A third model, referred to in the video as GPT-6 Astra, read the prompt and the resulting app and returned a pass or fail judgement out of 10. That reviewer is a second AI model, not a human tester, and the presenter's own hands-on verdict matched its call in every project.
| Project | Opus 5.5 score | Gemini 4 Argon score | Presenter's verdict |
|---|---|---|---|
| Flood simulator | 7/10 | 5/10 | Opus, despite a missing weir |
| Canva clone | 8/10 | 6/10 | Opus, much closer to Canva |
| Bike shop site | 8/10 | 5/10 | Opus, clear margin |
| Driving simulator | 6/10 | 5/10 | Opus, with a flicker bug |
| Kitchen designer | Not scored on camera | Not scored on camera | Opus, more detailed |
| Launch film | 7/10 | 4/10 | Opus, better storytelling |
Scores clustered tightly in two projects, the driving simulator and the flood simulator, where a single failed requirement could have flipped the result. The kitchen designer was compared on camera without a score being stated, so it sits outside the judged tally.
What the six builds revealed about each model's strengths
Claude 5.5 won on front-end craft and narrative structure, while Gemini 4 Argon matched it on simulation physics and produced the strongest raw 3D asset of the set. The pattern showed up in specific builds rather than as a general coding gap.
Flood simulator
Both models built a three.js and TypeScript flood scene of a river town in Ireland, and both barriers held water in a 100-year surge test. Gemini 4 Argon's version opened with buildings already underwater and a barrier script that would not stop running; Claude 5.5's version rendered so slowly that the presenter had it patched before testing. Opus also deleted a weir that exists on the real map, which the presenter called a fail.
Canva clone
Claude 5.5 opened on a template collection page and allowed in-place editing, format resizing and background removal, matching the Canva workflow the brief asked for. The background removal step depends on a local tool, and the presenter ran it without describing which model or service performed it. Gemini 4 Argon met the resizing requirement but buried text editing in a side panel with coordinate readouts.
Bike shop and booking engine
Gemini 4 Argon produced a technical 3D hero and a busy booking form. Claude 5.5 produced a product page, size selection, a reservation flow and a progressive booking wizard, which the presenter described as in a different league.
Driving simulator and launch film
Claude 5.5 had actual roads, collision detection against buildings, screen mirroring for infractions and a per-route fault report. Gemini 4 Argon had no road surface, no collision detection and buildings that could be driven through. For the 60-second launch film, Opus used scene transitions and a tagline; Gemini 4 Argon used small type, basic Notion and a vertical version that was harder to read.
Kitchen designer
Gemini 4 Argon delivered four photorealistic viewpoints per style, a fly-through and a three-style comparison, and the presenter called the Blender model great. Claude 5.5 returned more interior detail, including wall tiles and props, a PDF plan download and a parts list with castings.
Why run time and token counts complicate the comparison
Run times did not track quality in a straight line, and the video's own presenter attributed much of the token gap to a platform limit rather than model behavior. Gemini 4 Argon hit a five-hour session cap during two projects and restarted, which the presenter believes broke prompt caching and counted cached input as new input.
The flood simulator numbers are the clearest example. Gemini 4 Argon took 74 minutes against Claude 5.5's 54, yet Opus used roughly 2.5 times the total tokens because most of its extra volume was cached input. Output token totals were close. The presenter read that as Opus doing more work in less wall-clock time, which is a plausible reading of cached versus uncached input but not a controlled measurement.
Those tokens were consumed through a paid plan. The presenter signed up to a $99 monthly Google AI Ultra plan to access Gemini 4 Argon in Google Antigravity, and the video gives no per-token cost comparison between the two models. Any claim that one model is cheaper per finished project would need accounting for cache hits, retries and the five-hour restart, none of which the test isolated.
Treat the timings as wall-clock records of a single unattended run, not benchmarks. A rerun on different hardware, network conditions or cache state could shift them by tens of minutes, and the transcript names the presenter's RTX 5090 only in passing during the flood simulation.
Where each model fits in a real workflow
Pick Claude 5.5 when the deliverable is a client-facing interface or a story that has to look finished, and pick Gemini 4 Argon when the work is a physics or 3D asset where its output already held up. The test never covered refactoring, test writing, large codebases or multi-day iteration, so this is a guide for one-shot generation only.
- Choose Claude 5.5 for e-commerce pages, landing pages, booking wizards and any front-end where typography and layout hierarchy are part of the brief.
- Choose Gemini 4 Argon for Blender asset generation and simulation scenes where the underlying model or physics behaved correctly on the first pass.
- Budget for restarts if you rely on a session-capped plan, and keep the number of prompts in one session low enough to finish before the cap.
- Verify with a checklist, not a score. The GPT-6 Astra review missed a visible rendering flicker in Opus's driving simulator, which the presenter caught only by playing the build.
One caution applies to both models. The presenter described the front-end quality of the Claude models as strong, and separately said neither video output was impressive enough to ship without component libraries for animation. Skilled review of generated work remains part of the workflow.
FAQ
- Did Gemini 4 Argon beat Claude 5.5 on any build?
No. Claude 5.5 took the higher judged score in every project where a score was announced, and the presenter agreed in each case. The closest result was the driving simulator at 6 to 5.
- Who judged the builds?
A third model, called GPT-6 Astra in the video, read each prompt and build and returned a score out of 10. It is an AI reviewer, not an independent human study, and it missed a rendering bug that the presenter saw immediately.
- How much did it cost to run Gemini 4 Argon?
The presenter used a $99 monthly Google AI Ultra plan to reach Gemini 4 Argon inside Google Antigravity. The video does not report a per-token cost for either model, and restart-related token inflation makes direct cost comparison unreliable.
- Was Opus 5.5 faster than Gemini 4 Argon?
Yes on most projects. Gemini 4 Argon took 74 minutes against 54 on the flood simulator, and Opus finished more spec items per run, but a five-hour session cap and restart distorted the inputs for Gemini 4 Argon in two projects.
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