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

The author notes Gemini 4 Argon scores 97.5% on ARC-AGI-3 with its official harness but only 49.8% with the standard one, at higher cost, stressing how much harness choice affects results.

A post claims Gemini 4 Carbon has opened EAP, is public-testing on LMArena, beats Argon on 3D tasks but lags on frontend while following instructions better.

A user reports trying Gemini 4 in AntiGravity, praising its sharp document review and text ability rivaling Claude, while noting one task burned about 5% of the weekly quota.

Google AI Studio staffer Logan Kilpatrick says Argon handled roughly 25 trillion tokens in its first three days, far exceeding expectations.

Sundar Pichai says Argon used almost 25 trillion tokens in its first 72 hours, about 3x Gemini 3.0, and that Google is capacity constrained while serving it.

A Chinese post wonders why a new Gemini 4 Carbon checkpoint is surfacing so soon while the previously teased Gemini 4 Flash has yet to appear.

LuminaBench had Gemini 4 Carbon draw an Xbox controller as SVG with no tracing or search allowed, showcasing its vector output.

A Chinese user complains that Gemini 4 Argon's usage quota is far too small and that the promised Gemini 4 Flash release still hasn't appeared.

LuminaBench runs the SVG pelican-on-bicycle prompt on Gemini 4 Carbon, noting a mediocre UI and very slow speed but solid results across many tasks.

Google Poland's official account posted a Gemini 4 Argon demo video claiming ~3x faster Gemma 4 training, 2.3x faster xTB molecule prediction, and 0.7% attack success on Gray Swan IPI.

Early partners reportedly have Gemini 4 Carbon EAP, showing clear gains over Argon but slower speed and weaker frontend, now being tested in Arena.

A video compares Gemini 4 Carbon's voxel pagoda output against several frontier models, showing higher quality than Argon at noticeably slower speed.