Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

A user shares a koi pond video generated by Gemini 4 Argon, saying it took 3 minutes and $3, cheaper and faster than Opus 5.5, with the model at #5 on Arena's agentic code leaderboard.

Gemini 4 Argon, paired with Nano Banana in a multi-round self-iteration loop, procedurally generated a futuristic three.js 3D battle helicopter from scratch.

The same prompt asked Gemini 4 and Astra 6 to each build a playable 3D Flappy Bird in a single HTML file, with a video comparing the results.

The poster says Gemini 4 Argon finally beats Gemini 3.1 and 3.8 by a wide margin and reaches Opus 5.5 level, with a comparison video against 3.1 Pro.

The author says Argon produced a complete 3D medieval town in the browser from a few thousand lines of code, with no model files or image textures.

jjschnyder had Gemini 4 and Astra each play the Skyrim tutorial, calling both top-tier at understanding 3D, physics and real-time environments.

Argon completed the BridgeBench lamp test in under two minutes at a cost of $0.18, compared with Opus 5.5's 9 minutes 35 seconds and $1.27.

A Russian-language post says Gemini 4 Argon autonomously built a photorealistic pagoda and temple complex in Blender from a single prompt with no human intervention, using multiple AI agents to critique and refine its own work.

One user complained that trying Gemini 4 Argon forces an Ultra-tier subscription, reflecting access limited to the higher-priced plan.

Gemini 4 Argon scores about twice as high as 3.8-flash on Eyebench to take 5th place, though the poster is surprised that an Anthropic model beats Google's on a vision benchmark.

A developer says they are using Gemini 4 Argon to create CAD designs, without giving specific examples or results.

A commenter says Argon's creative writing is the most striking part, while criticizing Astra's creative writing and Opus 5.5 for taking roundabout routes due to heavy censorship.