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

The author says Opus 5.5 still tops Terminal-Bench, but Gemini 4 Argon felt ahead on large refactors and long-context work.

A user argues Argon could be a world model, building explorable 3D worlds purely in Blender code without any image, video or 3D asset generators.

Given the same prompt to animate a man throwing a raw turkey at someone, Gemini 4 Argon and Claude Opus 5.5 take different approaches, and the author asks which wins.

A user marvels at DeepMind's engineering, saying Gemini 4 Argon is now catching up with Opus 5.5 in practice.

An anime-style parody dramatizes Gemini 4 Argon's million-token output, DeepSWE 77.9%, autonomous bug fixing and pricing, ending on the joke that regular users still can't access it.

This AI Agents Pulse roundup notes Google made Gemini 4 Argon globally available via API in AI Studio and the Gemini platform, with AI Pro/Ultra subscribers also getting it in the Gemini app and Search AI Mode.

A user reports that a single Gemini 4 Argon prompt in the Gemini app consumed 27% of their 5-hour quota, calling the app nearly unusable.

Given the same prompt, Gemini 4 Argon cost $11 and took 15 minutes while DeepSeek V4 cost $1.2 and took 70 minutes; the author says DeepSeek won on price and design taste.

A new benchmark had nine models write C++ StarCraft bots in one hour offline; GPT-6 Astra led at 57, Gemini 4 Argon third at 39.

A developer demos several typography animation effects built with Gemini 4 Argon and shares the animation file plus skill.

A user calls a benchmark rigged for placing muse wonder third above Astra and Gemini 4 Argon.

The author mocks freeloaders who use Gemini 4 Argon while complaining, noting a 30 RMB monthly plan offers Argon in Antigravity plus YouTube Premium and 5TB cloud storage.