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

A developer used Gemini 4 Argon to improve vLLM code, lifting Gemma4 E2B throughput on A100 by almost 3x (up to 108k t/s) without hurting token consistency or quality.

The new AA Intelligence Index v5 suggests some open-source models were benchmaxxing, widening the gap with OpenAI, Anthropic and Gemini Argon.

A video post mocks a comparison between Gemini 4 Argon Fable 5.1 and GPT 6.1 Terra, saying they'd rather pretend they didn't see it.

A creator reimagined a one-vs-six fight clip originally made with Seedance 2.5 into three cinematic parts using Gemini Argon.

On Cognition's new FrontierCode 2.0 coding benchmark, Gemini 4 Argon scores 26.3% for 11th place, below Claude Haiku 5.5's 27.7% and far behind Opus 5.5's 45.5% and GPT-6 Astra's 42.8%.

A rumor says Gemini 4 Flash runs at ~300 tok/s while matching Gemini 4 Argon quality, which the poster says would kill long reasoning waits if priced like Flash.

A user claims GPT-6.1 Terra beat Gemini 4 Argon on the Artificial Analysis Intelligence Index and jokes that the AA team should get their eyes checked, hinting doubt about the ranking.

A post lists GPT-6.1 Sol, Opus 5.5, Gemini 4 Argon, Grok 4.7 and more, arguing the interesting question is what people build once model cost stops mattering.

A user reports Gemini 4 Argon is now accessible in Antigravity IDE for some accounts, and asks Google for quota resets and an early Gemini 4 Flash release.

A blogger says the Gemini 4 Argon web app exhausts quota after about two exchanges even on Pro, so they route it through Magpie and Antigravity into Open WebUI inside a Grok Bot VM for multi-turn chats.

A user shares a new AA index screenshot showing GPT-6.1 Terra ahead of Fable 5.1 and Gemini 4 Argon, questioning whether a $1 model ranking so high means the index is off.

A user asks whether their boss is right that Gemini 4 Argon is actually better than Opus 5.5 or GPT-6.1 Terra.