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

A Japanese user used Gemini 4 Argon to turn sales data and proposals into animated, clickable slides where charts open details and backgrounds move.

DeepMind showed four Gemini 4 Argon agents sharing one world, writing 2,783 programs via code and text only, and killing the Ender Dragon after 27 hours.

A user reports Gemini 4 Argon solved a player re-ID task zero-shot, handling occlusion and spatial tracking, and sees it as fit for automated analytics pipelines.

The creator photogrammetry-scanned a deer taxidermy into a web-optimized 3D model and animated it walking a 3DGS mountain path, combining Codex 6.1 Sol, Gemini 4 Argon, Claude Opus 5.5 and manual work.

A user finds Argon trades blows with Fable and Astra but excels at conversational balance, and predicts a Gemini 4 Flash within three months.

A user built a personal portfolio site with Gemini 4 Argon as the main model and 3.8 Flash as sub-model, spending only 7% of quota versus 20%+ before.

A commentary argues Gemini 4 Argon is strong enough to sway Claude veterans, but its quota burn pushes users toward Ultra, making pricing its biggest competitor.

An image comparing the usage allowances of Gemini 4 Argon, Opus 5.5 and GPT-6 Astra on the Pro subscription.

Another image post says Gemini 4 Argon has landed in Google AI Studio for both paid and free Pro users, slightly differing from other access claims.

A user reports Google granted AI Pro subscribers access to Gemini 4 Argon in AI Studio, testing 50 Playground runs and 2 Build runs, with the model marked Paid.

A Japanese roundup splits OpenCode's free and paid models, placing Gemini 4 Argon in the paid tier with GPT-6 and Opus 5.5 while GLM-5.3 Flash stays free.

A post lists per-million-token input/output API prices for Opus 5.5, GPT-6 Astra, Gemini 4 Argon and others while promoting the SayGM gateway.