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

A post lists five free guide links, the first covering best use cases for Google Gemini 4 Argon, plus image/video prompting, Claude plugins and custom GPTs.

A Japanese user shares an Umaibo-style illustration generated with Gemini 4.0 Argon, showing its image output.

A user says they are trying Gemini 4 Argon in paid Google AI Pro for the first time, testing limits with low and medium effort and planning to report back.

A meme post jokes that Gemini 4 Argon is claiming AGI to secure a $10B funding round, poking fun at the model's launch narrative.

A user posts an image asking whether Gemini Argon was supposed to be better than Claude Opus 5.5, implying results fell short of the claim.

A post says Gemini 4 Argon beat Minecraft's Ender Dragon after 27 hours of autonomous play, arguing the cost of intelligence now matters as much as the intelligence.

Google claims Gemini 4 Argon converted a single tabletop photo into a sim twin and 4,500+ episodes, raising real robot success from 56% to 75%.

A leak says Carbon is not a standalone model but an internal Gemini 4 checkpoint tested on Google's Jetski coding platform, coding near Opus 5.5 and not replacing Argon.

The post says Google is reportedly testing an internal checkpoint called Carbon whose coding feels close to Claude Opus 5.5, and that Argon itself came from an older Barium-B checkpoint, with no public benchmark yet.

A Google DeepMind paper reports four Gemini 4 Argon agents beat the Ender Dragon in 27h14m with no human help or guide, dying 605 times and using ~1.35B input tokens at ~$1,000 cost.

A short post asking whether anyone is still using Gemini 4 Argon, hinting at questions over its availability and traction.

The post recaps four Gemini 4 Argon agents beating Minecraft with zero human help: 2,783 self-written programs, 605 deaths and 27 hours, arguing agentic AI is the real trend.