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

A developer uses Gemini 4 Argon to coordinate agy CLI running gemini 3.8 flash for coding and codex CLI with gpt-6-astra-xhigh for review, auto-fixing small issues.

A user reports Arena went down as many tried to access Gemini 4 Argon through it, failing on a third attempt and wondering if a new checkpoint launched.

A user tested Argon High with an old vibe-check prompt: car shape is close to reference and speed is impressive, but colors look washed out, textures float, and the garage/UI feel dated.

A comment argues Argon is a good model but drains usage limits too quickly, and hopes Google ships the reportedly better Carbon with limits that last.

A user shares another Gemini 4 Carbon output, a Three.js piece generated in Arena.

A user posted a screenshot of a Gemini gray-test model, suspected to be gemini-4-flash, without further details.

Google AI Pro trial users reportedly see both Opus 5.5 and Gemini 4 Argon in Antigravity, with demos including 3D wizard chess, a periodic table, and a WebGL fur engine.

In the video, Wiz shows Gemini 4 pulling a master key from a health-software config file and turning it into admin access, framed as a dual-use AI security risk.

A post claims Google Research open-sourced a self-improving harness where Gemini 4 Argon plans architecture while 3.8 Flash writes code, lifting benchmark accuracy 34%.

Users report that Google AI Pro plans obtained via Jio are seeing a staged rollout of Gemini 4 Argon and Claude Opus 5.5, though official docs are unchanged.

A joke post notes Google's model codenames Argon and Carbon read like chemistry terms an intern picked at random.

A post claims leaked November model intel says Google's Gemini 4 Carbon matches Opus 5.5 in coding on the internal Jetski platform, hinting at recursive self-improvement.