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

Highlights from a one-hour Pichai interview: Gemini 4 Argon processed 25 trillion tokens in 72 hours, about 3x Gemini 3.0, is capacity constrained with distilled versions in the works, and reportedly beat Minecraft solo.

A short video plays on the 'Gemini 4 Argon comeback is personal' joke while showing the model's output.

A timeline post traces Argon's Sep 30 launch, internal Carbon (Jetski) testing from Oct 9 and partner/LM Arena access on Oct 21–23, guessing Carbon rolls out small as soon as next week.

The poster shares a screenshot of Gemini 4 Argon's answer to a question about its philosophy on life, showing how the model handles open-ended philosophical prompts.

A leak claims Google is dogfooding Gemini 4 Carbon internally and testing a Gemini 4 Argon Max mode plus a Carbon G medium reasoning tier.

An Arena comparison pits Kimi K3.1 Pro against Gemini 4 Carbon (Arena Checkpoint), Gemini 4 Argon and GPT-6 Astra, with the poster arguing Kimi K3.1 now genuinely competes with frontier models and beat Astra.

An image post claims Chinese model Kimi K3.1 Pro outperforms Claude Opus 5.5 and Gemini 4 Argon on all benchmarks, without showing numbers.

A user built a 3D airplane with Gemini 4 Carbon in Arena and found the lighting, detail and overall look surprisingly good.

A video runs the same one-shot prompt through Gemini Argon 4, Claude Opus 5.5, GPT Astra and Grok 4.6 and shows the differing outputs for viewers to judge.

Leaked identifiers show gemini-4-carbon and gemini-4-argon-max-mode now in Google internal teamfood testing.

A tip for using Gemini 4 Argon to export an image's colors, typography and composition as reusable JSON style data.

A Japanese AI chat tier list rates models by conversational feel, placing Gemini Argon in the 'unmanageable extrovert' tier alongside Fable5, Opus5.5 and GPT6.