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

A user reports that asking Gemini 4 Argon to print a single sentence burned 2% of the 5-hour usage limit.

DeepMind's Genie 4 world model demo has Gemini 3.8 Flash break tasks into steps and replan, plus an experiment where Gemini 4 Argon drives an Apollo humanoid playing checkers.

Commentary argues that if Gemini 4 Flash delivers a big speed jump without hurting reasoning and coding, the upgrade could be larger than expected.

In a multiplayer autoresearch run, Gemini 4 Argon set each frontier and rose from 0.625 to 0.938 in round 3 by citing earlier failed attempts; Fable and Sol then built on its strategy.

A screenshot post showing Gemini 4 Argon replying that humans possess a superpower, shared as a fun output sample.

A team combines Gemini 4 Argon with Genie 4 for robotics: Argon splits user instructions into steps fed as live captions to steer Genie 4's simulation as a task progresses.

Google says it paired interactive world model Genie 4 with Gemini 4 Argon in DeepMind robotics research to simulate scenarios in imagined worlds.

Google's Gemini team is reportedly naming Gemini 4 models after periodic-table elements: Argon, Carbon (rumored next Gemini 4 checkpoint) and Hydrogen (rumored Gemini 4 Flash).

The poster claims the insanely fast Gemini 4 Argon seen in Arena testing is actually the "hydrogen bomb" Gemini 4 Flash at 300tps.

An image post lays out a possible Google Gemini lineup—Argon, current Pro-tier Carbon, a heavier internal checkpoint Hydrogen, and the Flash most people would run—while noting all of it is still rumor.

A single HTML file with procedural Three.js renders a 7.75M-voxel scene with a pagoda, bridge, viaduct and steam train at 60fps for $2.48.

A leak claims Gemini 4 Flash runs around 300 tps, faster than GPT-6.1 Terra, with Argon-level intelligence and pricing below Argon, near Sonnet 5.5.