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

The author used Aprilzz423's skill with Gemini 4 Argon to build a small F1 pit stop game that ships with a hand-drawn wobble slider and color toggles out of the box.

A Turkish weekly recap says Gemini 4 Argon is now public: available as gemini-4-argon in the API, in the Gemini app for Pro and Ultra subscribers with a 1M-token input/output window, and in the Antigravity coding agent for Ultra and paid Pro only, rolling out gradually.

A poll asks readers to pick the biggest token eater among Opus 5.5, Grok 4.7, GPT-6 Astra and Gemini Argon 4.

The author waited for Argon to compare UI design output against Opus 5.5, but still finds Argon behind and jokes about waiting for Carbon instead.

A demo video on canyon radar blind spots and Bayesian inverse tracking was rendered with 3D content produced 100% by Gemini 4 Argon.

In a same-prompt one-shot test building a 1996 Paris scene in pure code, Gemini 4 Argon cost about $2.50 and 30 minutes, versus $73.80 and nearly 3 hours for Opus 5.5 and $0.34 for DeepSeek Flash 1017.

A video compares Gemini 4 Argon and Opus 5.5 on one prompt — a molten elephant emerging from a carpet — with Argon rendering a lava boss while Opus just sets the elephant on fire.

An opinion post argues Claude Opus 5.5 suits complex work, GPT-6 Astra specialized tasks, Gemini 4 Argon value and factual reliability, and DeepSeek V4.1 Flash low-cost open weights.

On a simple task, Argon made 142 requests, compacted context twice, exceeded 800K context and burned millions of tokens, which the author calls terrible work habits.

A user relays a teaser that Gemini 4 Flash is coming soon, without a date or official source.

Free Google AI Pro subscriptions from student offers, Jio bundles or Pixel promos don't get Argon in Antigravity, but the author says Gemini 4 Argon is still usable through Google AI Studio.

In a three-arm, six-layer tower benchmark, Robotics-ER 2 finished 6/6 layers in 30.63s while Gemini 4 Argon stalled at 5/6 after about 16s, exposing workspace-geometry gating issues.