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

patloeber highlights Luciano Martins' work using Gemini 4 Argon to speed up vLLM and Gemma, a developer-side performance effort.

A new Google DeepMind paper reports four Gemini 4 Argon coding agents self-organized with no human help, wrote 2,783 programs and killed the Ender Dragon in 27 hours 14 minutes.

A post lists per-million-token input/output pricing for GPT-6 Astra, Opus 5.5, Gemini 4 Argon and others, with Argon marked as a limited-time $2/$10.

A reply post that only says "According to Gemini Argon:" with an attached image, giving no context on the question or conclusion.

The author pairs Gemini 4 Argon with the transkun piano transcription model to analyze and visualize their playing, separating left and right hands from MIDI and video.

AI Studio Build works well with Gemini 4 Argon and adds one-click UI chips for making games, building Three.js 3D scenes and using DeepMind's science skills.

An image post shows a take from Gemini 4 Argon that the poster describes as crazy, with the details visible only in the image.

A short post saying only "beautiful stuff !! Gemini 4 Argon" with no details or evidence.

Android Bench added Claude Sonnet 5.5, Gemini 4 Argon and GPT 6.1 Sol, producing a new #1 on the leaderboard for model selection.

Respan AI Gateway's week-two discount list includes Gemini 4 Argon and Gemini 3.8 Flash at 50% off, alongside a 40% cut on GPT-6.1 Sol.

A user calls GPT-6.1 Terra's AA Intelligence Index score of 48 disappointing and says they are now waiting for Fable 5.5 and Gemini 4 Argon.

Ortet's new evaluation puts Gemini 4 Argon first overall on its Health Index and first in biological research, with Gemma 4 also standing out.