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

Two videos use the same prompt to have Gemini 4 Argon and Muse Wonder rebuild and animate Transformers' Bumblebee in three.js, reproducible via the atomic.chat API.

Hyperbrowser now supports Gemini 4 Argon, letting it browse, click and type autonomously to complete multi-step tasks in Gemini Computer Use.

Guests on the thursdai podcast talk at 1:30:36 about what Gemini 4 Argon means for DeepMind and why they are excited about it.

Researchers used Gemini 4 Argon to find 3D magnetic plasma solutions and draft paper sections and verification code, with results independently checked; the work concerns plasma confinement math related to fusion, not fusion itself.

Argon is said to run tuning without anyone at the keyboard, cutting long session prep by 40% and lifting Gemma 4 E2B from about 37,900 to over 108,000 tokens per second on an A100.

A user shares a video showing Gemini 4 Argon able to recreate a task while GPT-6 Astra fails on the same task.

A user says Gemini 4 Argon scores very well on most benchmarks but performs poorly on reliability, hallucination and vibes-style evaluations.

The author tested Gemini Argon, Opus 5.5, Sonnet 5.5 and Gemini 3.8 Flash on AI video generation with the same single-shot prompt, and will share the results later.

The author posted a three-way video comparison including Gemini 4 Argon and says the winner is obvious at a glance.

An AI YouTuber celebrating 200 subscribers uploaded a brief hands-on overview video of Gemini 4 Argon.

After a full day of testing, the user says Argon is a big upgrade over older Gemini models and faster than Opus 5.5, but hallucinates often, is expensive, excels at 3D work like Three.js and Blender, and is still less trusted than Opus 5.5 for daily coding.

A developer says they only realized how good Gemini 4.0 Argon is after forgetting to switch models and doing a task with Gemini 3.8 Flash.