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

A reply shows a user asking Gemini to update a graph labeled with Argon's high setting, touching on Gemini 4 Argon tier info.

Google says Gemini 4 Argon lifted gemma-4-E2B-it from ~37,900 to 108,000+ tokens/sec on an A100 via vLLM with 100% token parity and no quality drop, across 8 PRs.

After two days of trying Gemini Argon as a workflow advisor and failing, the author argues it lacks big-model knowledge density and is likely a small Flash-tier model, with larger releases to come.

A weekly roundup says Kimi K3.1 Pro matches Claude Opus 5.5 and GPT-6 Astra on several benchmarks and on Arena agentic tasks trails only Gemini 4 Argon and Muse Wonder Full.

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.