Hands-on pits Gemini 4 Argon against Claude Opus 5.5 and GPT-6 Astra on identical prompts

The test

On October 13, 2026, Google made Gemini 4 Argon generally available. The next day, a hands-on review published on dev.to by Anass Kartit put the new model against Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra using identical prompts: a 3D Lamborghini built entirely in code, a one-shot 3D kart game, an animated SVG drawing, and 82 short labelled questions. Argon ran in opencode, Opus 5.5 in Claude Code and Astra in Codex, so as the author notes, the headline comparison measures agent setups rather than bare models, and each test was run only once.

A 3D car from one image

The showpiece was a Blender build. After Codex produced a single reference sheet of a supercar seen from the front, side, back and top, Argon, connected to Blender through an MCP server, modelled the body, wheels, wing and glass part by part in Python, ending with 43 objects, working headlights and taillights, a turntable rig and a 240-frame camera move.

That session cost $22.13 on the author's Gemini API key: 61 steps, just over 10.2 million input tokens and about 127,000 output tokens, with only around 4% of input served from cache. Because every step resent the entire conversation including images, the author advises keeping agent sessions short and keeping prompt prefixes stable, since Google's implicit caching applies only to shared prefixes of 6,144 tokens or more.

Three agents, one Lamborghini

The main coding test asked for a three.js supercar written entirely in code, with no model files, driving down an endless neon highway with a chase camera, speedometer, turbo exhaust flames, six paint colours and an orbit mode. According to the dev.to write-up, Astra in Codex finished in 9 minutes 17 seconds with 193 lines, Opus 5.5 in Claude Code took 14 minutes 39 seconds for 1,166 lines, and Argon produced its first file in about 7 minutes and then spent another 17 minutes self-testing on the way to 2,540 lines, at a cost of $2.45 on an API key. The subscription runs for Claude Code and Codex were converted to roughly $14.14 and $3.30 at OpenRouter list prices.

The author's verdict on looks went to Argon, citing the wedge silhouette, Y-shaped lights and rear wing, while calling its driving feel the weakest of the three.

One-shot calls and short answers

On a single API call through OpenRouter to generate a working 3D kart game, Argon was both fastest and cheapest: 220.6 seconds and $0.306, scoring 9.5 out of 10. Astra needed 323.4 seconds and $0.704 but delivered a 10 out of 10 result, while Opus 5.5, at 307.6 seconds and $0.641, was cut off at the 32,000-token limit, with 20,770 of those tokens spent on reasoning, and produced nothing usable. On the SVG drawing task, Argon at high thinking produced the most detailed car of the three.

On 82 labelled questions expecting one exact word, Argon and Astra both answered 82 of 82, and Opus 5.5 scored 81 of 82, its single miss being a formatting slip rather than a logic error. Argon was the slowest here, with a 3.355-second median against 1.701 seconds for Astra.

Configuration friction

Argon did not run in opencode on the first attempt, and the review documents two fixes. Argon can think for more than five minutes before emitting its first token, tripping opencode's default header timeout, so the author raised headerTimeout and chunkTimeout to 1,800,000 milliseconds. Separately, opencode 1.19.0 caps replies at 32,000 output tokens, and three runs ended at exactly that limit with the file silently dropped; setting OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX=65536 resolved it. In opencode v2, the author notes, the cap follows the model's own output limit instead, and Google's default output limit for Argon is 65,536 tokens.

What Google itself reports

Google's launch table, quoted in the review, puts Argon ahead on agentic coding and long context, at 77.9% on DeepSWE v1.1 and 84.2% on GraphWalks 256k-1M, but behind on Terminal-bench 4.0, 57.4% versus 66.4% for Opus 5.5, and on FrontierSWE v2, 55.0% versus 65.5% for Astra. These are vendor-reported figures, and the review's own tests carry the usual caveat that model outputs vary from run to run.

Why it matters

Argon ships with a 1,048,576-token context window and multimodal input, priced at $2 per million input tokens and $10 per million output tokens until November 30, 2026, doubling to $4 and $20 afterwards. The dev.to tests suggest it is currently the strongest of the three for visual output and competitive on cost and speed in one-shot generation, but not a clean winner everywhere, and it came with real operational friction. Two practical lessons stand out for anyone running agents: token bills compound quickly when an agent resends a multimodal context at every step, and tool-side output caps can silently discard long single-file replies. The choice between frontier models still looks task-dependent, with Astra favouring polished one-shot apps, Opus 5.5 offering the zero-configuration experience, and Argon leading on visuals.