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

GIGAZINE reports four Gemini 4 Argon AI agents cleared Minecraft in 27 hours 14 minutes without human assistance, with the poster noting the 165,000 yen cost.

A developer shares that a small project built with Gemini 4 Argon was praised by Derya, noting Argon lets you build many impressive things.

A rundown of Google AI Pro quotas: Gemini 4 Argon and 3.1 Pro get higher usage limits, alongside Deep Research, Jules, Flow and other allowances.

Someone used Gemini 4 Argon to build a Subway Surfers-style game featuring a flying Tesla Roadster, with traffic dodging, aerial flips and power-ups; the game and code are released.

A meme pits Fable 5.1 and GPT-6 Astra against Gemini 4 Argon, joking that Argon is the genuinely costly one.

A video compares Claude Opus 5.5 and Gemini 4 Argon building the same island village brief in Blender and three.js.

The user says they never got a GeminiApp trusted-tester onboarding email but already tested Gemini 4 Argon in early access on Antigravity.

A demo shows Gemini 4 Argon in Antigravity building a 3D periodic table with orbiting electrons, atomic data and light spectra, synced to Tom Lehrer's Elements song.

A short take that Gemini has improved noticeably for autoresearch and recursive self-improvement tasks with Argon.

Gemini 4 Argon leads Isoquant's new inference engineering benchmark with a 4.38× mean speedup across 9 tasks, ahead of GPT-6 Astra (3.71×) and Muse Wonder 1.0 (3.66×).

A user reports that Gemini 4 Argon's weekly limit shows 0% in Antigravity, quota resets on Oct 25 despite waiting since Oct 18, and AI Ultra on Gemini Web versus AI Pro in Antigravity may be conflicting.

TheZvi asks for reactions to Gemini 4 Argon and whether Google is back, saying a full post depends on the feedback received.