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

Given the same prompt, Opus 5.5 mapped a semiconductor fab in granular detail while Gemini 4 Argon produced a more compact layout.

A user reports that Gemini 4 Argon fully generated a Blender greenhouse scene with strong architecture and lighting, though the butterflies look like maxed-out abstract art.

Comparison numbers show Argon at 53 intelligence, 82 tokens/sec and ~$1.99 per task versus 3.8 Flash at 41, 136 tokens/sec and ~$1.24 per task.

A user had Gemini 4 Argon build a full butterfly conservatory in Blender; the greenhouse looks great but some butterflies came out deformed, shown in a video.

The author reruns the same prompt on the newly updated Gemini 4 Carbon and shows a before-and-after comparison in a video.

A Chinese poster compares Arena's Carbon checkpoint with Argon, saying Carbon is stronger on voxel detail, structure and intent, and citing rumors that its coding nears Opus 5.5.

A reviewer finds Gemini 4 Carbon beats Argon at 3D and game understanding but is slower and weaker at frontend, still behind Opus 5.5 and Astra at 3D.

Polymarket gives Gemini Carbon a 92% chance of being a Pro-tier model, likely Gemini 4, said to lead in 3D and be less lazy than Argon but still weak at coding and frontend.

A developer says testing Gemini 4 Argon via Devin's CLI and desktop app is ridiculously cheap, and now uses SWE-2 as a daily driver model.

A commenter argues neither Gemini 4 Argon nor Opus 5.5 generates video, so the effect is likely code rendered live in the browser.

The post claims Gemini 4 Argon agent swarms uncovered a 12-year-old 9.2-CVSS zero-day in core PHP (PHP2Shell / CVE-2026-109618), and the author built an interactive 3D attack-vs-exploit simulation in a single HTML file.

A developer demos MicroDuck, a home robot planned with Gemini 4 Argon and trained on DGX Spark, detecting falls without sending camera frames to the cloud.