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

A weekly AI roundup says Google's Gemini 4 agents coordinated without human help to beat Minecraft's Ender Dragon, alongside firms blocking AI agents.

The poster compares two Gemini 4 checkpoints, rates Carbon above Argon, and says they can't wait for release.

A leak says Gemini 4 Carbon is launching very soon, with closed testing underway and the model already listed on LM Arena.

A user jokes that after switching to the cheaper Gemini 4 Argon, they quietly reopened Claude two hours later.

A Korean post says Gemini 4 Argon handled 25 trillion tokens in its first 72 hours, three times faster than Gemini 3, straining Google's compute.

Rumors say Carbon is in partner early access and running on LMArena, stronger at 3D than Argon, slightly slower but less lazy, with a rough UI and a public launch nearing.

The author used Gemini 4 Argon with an SVG skill and a reference video to generate a consistent Pagoda animation from a single prompt.

A video made with Gemini 4 (Argon) illustrates how far AI video generation has come, framed as a shock to human video editors.

A user dropped Claude from 5x Max to Pro and upgraded to Gemini Ultra, letting Claude Code supervise while Gemini 4 Argon and Grok Build do the work.

A leak claims Gemini 4 Carbon brings meaningful upgrades over Gemini 4 Argon and could arrive in the next couple of weeks.

The poster claims Google TPUs handled roughly 25 trillion tokens in three days and calls Gemini 4 Argon a strong model.

A user shares screenshots calling Gemini 4 Argon the funniest AI model, showing its humorous outputs as a hands-on impression.