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

Theo blind-tested Argon on his T3 codebase, scoring it 3.95 with 51% false claims, and suggested developers are better off not using it.

Gemini Argon 4 (apparently Gemini 4 Pro) reportedly scores 12% on MazeBench, placing second just below GPT-6 Astra.

A user shares a video of Gemini 4 Argon's visual output, saying it isn't perfect but is clearly better than Opus 5.5 and leaves them impressed.

Google says Argon (Gemini 4) tackles complex coding and agent problems such as legacy Chromium code and Gemma speedups, and is taking applications for free high-quota access from researchers and cyber defenders.

An image post titled "cold truths from Gemini 4 Argon" keeps its specific claims inside the attached image.

A user says Gemini 4 Argon output a 330KB+ .md file of over 90k tokens in one go, and ranks it alongside Claude Opus 5.5 and GPT-6 Astra as one of the three strongest models.

A short review says Gemini 4 Argon carries on Google's strength of making codebases more modular, faster and better performance-optimized.

Google DeepMind released a video showing Gemini 4 Argon acting as a virtual builder in Antigravity to autonomously generate a fully explorable 3D pagoda scene in minutes.

The author used a single prompt with Gemini 4 Argon in Antigravity to build a drawer-wall climbing puzzle demo, published it on GitHub, and discussed how long it might take to reach AAA quality.

Roboflow says Gemini 4 Argon tops almost every vision evaluation and has begun testing it on segmentation, keypoints and video reasoning, areas where most VLMs are not yet usable.

Antseed showed Gemini 4 Argon combining images, video and motion under the same code, letting users keep a scene and direct the action inside it, with 10 generations costing $1.56 via Antseed.

A developer shows a business tycoon game that is 85% complete, built with Opus 5.5 and Sonnet 5.5, and says they have not tried Gemini 4 Argon yet but plan to.