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

A user ran the same prompt on Gemini 4 Argon and Opus, saying Argon trails Opus but is surprisingly strong for a Gemini model, with video attached.

Given the same brief, Opus 5.5 and Gemini 4 Argon each modeled a playable chip fab in Blender in 70 and 85 minutes, with different floor layouts.

The author ran Three.js projects with Gemini 4 Argon High inside Antigravity for three days and shares a prompt that requires verifier scripts first, numeric bands taken from reference images, and pixel-based measurement.

A short post asking for rankings, placing Opus 5.5 above Gemini 4 Argon and GPT 6 Astra.

Reports say Argon can drain a week's quota in five prompts, but this user still had 92% left after daily coding, PDFs and browser automation, and asks others to share their remaining quota.

A short post asking the DeepMind team to bring a particular personality or tone to the upcoming Gemini 4 Flash model.

A video demo shows Gemini 4 Argon in Antigravity producing browser-ready 3D work, including a real-time hair physics engine and a 3D periodic table.

The author had Opus 5.5, DeepSeek V4.1, MiniMax M3.1 Flash and Gemini 4 Argon each one-shot a rocket launch in pure code, comparing cost and time.

Phil Schmid explains that a single Gemini Managed Agents API call provides Python/Node/Git/Bash, network access, a credentials API and a persistent filesystem to run Gemini 4 Argon.

A user complains Antigravity's auto-task UI only says "Pro Model", making it impossible to tell whether Gemini 3.1 Pro or Gemini 4 Argon is running.

Kaggle and RunRehearsals released a public benchmark for predicting real human behavior; Gemini 4 Argon scored 80.6%, second behind Engram 1.0 at 94.1%.

A user predicts Gemini 4 Flash will mirror GPT-4o in 2024: not the smartest model, but fast, cheap and everywhere, and it will clear everything.