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

A post claims Gemini 4 Argon, running in Google's Antigravity multi-agent environment, helped physicists find new solutions overturning and extending the 59-year-old Grad fusion conjecture within hours, linking a TechRadar report.

The author tried a mechanical bee test on Gemini 4 Carbon in Arena but got Mimo-v2.6-pro instead, and the bee's flight broke down hilariously on video.

Gemini 4 Carbon is reportedly showing up on Arena under Gemini 4 Argon, with initial Xbox controller tests feeling better than Argon while more testing continues.

Shivam posted a video comparing Gemini Argon and Claude Opus 5.5 on video generation, with no stated conclusion or scores in the text.

A user cites Gemini 4 Argon usage numbers as evidence of strong demand and asks whether Google tracks how much comes from humans versus agents talking to each other.

A video compares Gemini 4 Carbon's voxel pagoda output against Claude Opus and Sonnet, finding Carbon more detailed in structure but weaker in background elements than Opus.

A video demo praises Gemini 4 Carbon's 3D output as a huge leap, predicting public access before December.

A user claims Gemini 4 Carbon is imminent with even lower TPS, complaining that Google is moving on before Argon users get comfortable.

The author says Opus 5.5 still leads SWE-bench and Terminal-Bench, while Argon wins DeepSWE and 1M-context refactors, favoring Argon for large codebases.

LuminaBench says Gemini 4 Carbon checkpoint 1 is in Arena testing, producing its cleanest Super Mario game output in one try at roughly 30 minutes, and no longer lazy like Argon.

Given the same prompt to build a chip fab in Blender, Claude Opus 5.5 and Gemini 4 Argon produced different layouts and timings (70 vs 85 minutes), though the author cautions it isn't real fab simulation.

Running the same concurrent LFU cache prompt in two separate chats consumed 25% and 53% of the 5-hour quota, with Low thinking offering little savings.