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

The poster suspects Google and OpenAI hand creators free Ultra subscriptions just to generate more 3D Gemini 4 Argon posts, while admitting they can't prove it.

Across 120 real robot runs, Gemini 4 Argon and Opus 5.5 each scored about 48%, GPT-6 Astra about 47%, while Kimi K3.1 Pro managed only 9%.

A rumor post says OpenAI resumed frontier training after safety upgrades, with GPT-6.1 Astra set to answer Opus 5.5 and Gemini 4 Carbon.

LuminaBench shares a blog.google link to an official post about Gemini 4 Argon aimed at builders.

A video claims Gemini Argon completed a task framed as bad news for Opus 5.5.

Video compares cost and time on one task: Kimi K3.1 at $2.04 in 8 minutes, Gemini 4 Argon at $6.10 in 9 minutes, with Opus 5.5 and GPT-6 Astra costing more.

A Tokyo nighttime helicopter-view video originally made with Seedance 2.5 was recreated by the author using Gemini Argon.

A roundup claims Kimi K3.1 Pro, MiniMax M3.1 Flash, Muse wonder and Gemini 4 Argon all flopped, leaving DeepSeek V4.1 Flash 1017 as the only model people praise while using.

A user claims Jio Pro, Edu Pro and even full-price Pro can be routed to a quantized gemini-4-argon-tiered pool, and that Ultra is needed for the full model.

A post relays Google's official blog demos for Gemini 4 Argon, including seaplane simulation, a live 3D Earth, an interactive jet engine and single-image 3D carriages.

A user used Google Antigravity with Gemini 4 Argon to turn a Kindle idle for five-plus years into a live token and task dashboard in half an hour.

A video compares Gemini 4 Pro and Claude Opus 5.5 on a 3D floatplane simulation, praising Gemini's physics and movement as impressive for a pre-release model.