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

Gemini 4 Argon is now live for Pro accounts on the Gemini App, with low, medium, and high reasoning effort options.

LexnLin says Argon is much better at 3D and decent at coding, but weak on frontend and motion with little taste, though overall a successful release.

A meme pokes fun at the typical AI-sounding output you get when asking Gemini 4 Argon to write a tweet.

A researcher says their team spent weeks working with another to improve Gemini's game dev capabilities, and congratulates the Gemini 4 Argon launch.

Someone tested Gemini 4 on self-built private science and medical benchmarks and called it impressive on real-world scientific and medical tasks, without sharing scores.

A team with early access to Gemini 4 Argon ran 250+ evals, scoring each cell against the cited filing figures, and found 99%+ matched, with frontier results on 22 of 25 hardest tasks.

A Chinese user shared a video saying Gemini 4 Argon used Fourier rotating circles to sketch a pelican riding a bike, showing its creative generation.

On the same Three.js 3D humanoid robot simulation task, the author says Gemini 4 Argon delivered and overtook Claude Opus 5.5 in several key areas.

Google published a Gemini 4 Argon migration checklist covering 1M token output, and Antigravity users can install the gemini-api-dev skill to migrate an app with one command.

Dan Shipper says Gemini 4 Argon scores only 21% on a senior engineer benchmark task, far below Sol 6 and even Muse.

An overseas author used Argon to render 534 artworks entirely in code and open-sourced the source, with Logan Kilpatrick liking the post, cited as evidence of native multimodality.

The poster says Gemini 4 Argon beats GPT-6.1 Sol on their benchmarks but does not sit on the Pareto curve, and offers it at 75% off until December 1, 2026.