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

The author pits Gemini 4 Carbon (updated checkpoint), Gemini 4 Argon, Kimi K3.1 and Qwen 4 against Opus 5.5 in Arena, betting Gemini comes closest.

The same prompt asked Gemini 4 Carbon and Grok 4.7 to build a 3D cow riding a cycle in three.js; the author mocks Grok's result and asks for Grok 4.8.

Google shows Gemini 4 Argon autonomously looping Blender MCP in Antigravity to build a Gothic cathedral, align camera grids and stage audio-reactive lighting from one Nano Banana image.

A student used Gemini 4 Argon to turn two course PDFs into podcast scripts, then listened via ElevenReader instead of attending class.

A user's subscription stack: Claude Pro for Opus 5.5 and Haiku 5.5, Google AI Pro for Gemini 3.8 Flash daily work plus Gemini 4 Argon medium for web deep research, with xAI, DeepSeek and ChatGPT Go filling gaps.

A leaker says the Gemini 4 Carbon checkpoint outperforms Argon but is unimpressive next to Opus 5.5, alongside takes on several other rumored models.

The author had Gemini 4 Argon study his posts, newsletters and YouTube speech to build three skill files, then write a post in his voice that he rated 9.2/10.

Hugging Face releases Carbon-X, an open 14B DNA model scoring 67% on its DNA benchmark versus 50% for Gemini 4 Argon.

The NSA warns of two PHP flaws dating back to 2004; Wiz's Atlas agent reportedly helped find them using Gemini 4 Argon and GPT-5.6 Sol.

A short post asks whether the hype around Gemini 4 Argon has already faded, reflecting community discussion of its release cadence.

A joking post noting friends still casually send Kimi K3.1 and Gemini 4 Argon test results while the world seems to fall apart.

A user showcases a Spider-Man mod for Minecraft built entirely by Gemini 4 Argon, animations included, calling the result insane.