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

A tester says Kimi K3.1 underperforms Opus 5.5, Astra and Gemini Argon on coding and costs more via subscription, contradicting its benchmarks.

fofrAI shows a code-art quine made by Gemini 4 Argon where running the code produces the next frame, depicting a pelican on a bike.

An interview with Pichai reportedly confirms Gemini 4 Argon processed 25 trillion tokens in 72 hours, 3x Gemini 3, forcing Google to cut compute elsewhere and skip 3.5 Pro.

Google DeepMind's Jon Barron found two numerical bugs in softmax and cross-entropy implementations, with Gemini 4 Argon writing the paper draft and figures and Flash 3.8 running most experiments.

Four top models built the same 3D whale-breaching scene; Kimi K3.1 cost $0.75, Gemini 4 Argon $2.30, GPT-6 Astra $2.00 and Claude Opus 5.5 $3.01.

A user animated their own piano rendition of "Caravan" using Gemini 4 Argon, demoing its audio-video generation.

Four models built the Eiffel Tower from one prompt: Kimi K3.1 at $2.04/8 min, Opus 5.5 at $8.95/10 min, Gemini 4 Argon at $6.10/9 min and GPT-6 Astra at $7.45/8 min, with Kimi judged most accurate.

Google lists Gemini 4 Argon's research impact: counterexamples in plasma physics, five lab-verified enzymes designed in hours, rare-disease genome analysis, and a quantum error-correction kernel with nearly double throughput.

A roundup of Gemini 4 Carbon rumors: an internal experimental branch alongside Argon focused on complex coding and long-running agentic tasks, with reported SWE-bench gains.

A user compares Gemini Argon with Claude, noting more frequent check-ins, fewer YOLO actions, slower trivial responses, but comparable overall competence.

The rumored Gemini Carbon shows up in LMSYS Arena with heavier reasoning and 2–4x slower speed than Argon, reportedly fixing Argon's reliability and long-horizon gaps.

A short post ranks the unreleased Gemini 4 Carbon above Kimi K3.1 Pro and Astra, saying it is waiting for the Carbon moment.