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

A user posted a long video testing the first output of Gemini 4 Carbon's second checkpoint, calling its interfaces and tools cleaner than Argon and faster than v1.

An infographic lists 71 companies training their own models, noting Gemini 4 Argon holds #1 on LMArena's text leaderboard on a preliminary score, alongside xAI's rename and DeepSeek's OpenRouter share.

A user complains that a single tweet consumed $4 worth of Gemini 4 Argon tokens, hinting at high usage cost.

A developer rates coding feel across Sol, Opus 5.5, Fable, Gemini 3.8 Flash and Gemini 4 Argon, finding Argon's Japanese natural but Pro-plan quota too tight to judge coding.

A user shares output from a new Gemini 4 Carbon checkpoint in Arena: a photorealistic 3D mechanical owl built with Three.js.

A developer used the second Arena checkpoint of Gemini 4 Carbon to generate a voxel-style 3D scene, reporting clearly better composition, color, lighting and reflections than Gemini 4 Argon.

LuminaBench says Gemini 4 Carbon generated a full FPS zombie game in one shot over about 30 minutes, making Argon look bad by comparison.

Opus 5.5, Sol 6.1, Gemini Argon and Grok 4.7 each built armies of soldier agents that could negotiate and ally; Gemini was the only one that actively lied.

Leaks from LMArena testers and DeepMind insiders say Gemini 4 Carbon makes a large leap in 3D scenes and spatial geometry, thinks slower and deeper, and was tested on Google's internal Jetski coding platform.

A PHP patch leaked two days early and the NSA issued an unusual advisory; the poster claims Gemini 4 Argon and an OpenAI system independently found the same vulnerability.

The poster had Gemini 4 Carbon recreate two reference images of black hole horizons entirely in three.js in 1h40m, and says Argon's attempt on the same task is coming.

A short Japanese post repeats "Gemini 4 Argon" with ellipses and question marks, hinting at a discovery without stating what it is.