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

On the 226-problem ErdosBench, DeepSeek v4.1 Flash 1017 ranks 6th, tying Gemini 4 Argon and Muse Wonder at a fraction of the cost.

A four-model mechanical peacock test pits Gemini 4 Carbon against Opus 5.5, GPT-6 Astra and GPT-6 Sol, with the author picking Carbon as most realistic.

A walkthrough notes Google AI Studio's free tier still offers gemini 3.8 flash, 3.6 flash, 2.5 pro/flash and gemma 4 with no card, while Gemini 4 Argon and most image generation are paid.

The author shares code and a live link for a Paris city demo built while experimenting with Gemini Argon, plus a few rougher extra cities.

A developer demos an endless 3D European street made with Gemini 4 Argon that starts as a hand-drawn ink sketch and blooms into watercolor as you walk or hover.

A roundup post claims Google gated Gemini 4 Argon through Project Fairwind because it surfaced real unpatched vulnerabilities, grouping it with OpenAI and Anthropic safety incidents.

A video shows Gemini 4 Argon inside Antigravity generating a single-file Three.js test in one shot, highlighting its front-end code generation.

A user says Gemini Argon matches Astra/Sol 6.1 overall and is substantially better for UI/UX work, while wishing Google offered free resets.

A video gives the same prompt to Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and Grok 4.7 to simulate a 9-cylinder radial engine, judging master-rod linkage, firing order and combustion timing.

A breakdown of Sundar Pichai's interview says Google prioritizes frontier research over Cloud revenue, with Gemini 4 Argon processing 25 trillion tokens in 72 hours.

A user says Google moved student accounts into a trial tier without Opus 5.5 or Gemini 4 Argon access, and upgrading led to a forced refund and loss of student status.

On Apple's RSI-Inference benchmark, where AI modifies vLLM/SGLang engines, Gemini 4 Argon speeds inference 1.39x and passes 81% of Day-0 hidden tests, though Pass@1 is only 6.7%.