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

LuminaBench posted a video showing Gemini 4 Argon drawing the Golden Gate Bridge using Fourier epicycles.

A video claims Argon reaches SOTA on real robot capabilities, tied with Opus 5.5 and above GPT-6 Astra, at $2.34 per attempt versus $4.96.

A user posted a video comparison claiming the free Gemini 4 Argon clearly outperforms Muse Wonder, without giving specific test conditions.

With Gemini 4 Argon officially released, Demis used it to build a wizard chess game demo with real physics and 3D modeling, marking the start of the Gemini 4 era.

Inspired by another creator's code-only 3D work, the author asked Gemini 4 Argon to build Leonardo da Vinci's inventions in 3D without Blender, and shared the results.

A tester compared models on pointing at objects in everyday photos with some scene understanding, finding Gemini 4 Argon surprisingly strong while noting trade-offs for Mistral Large 4, Claude Haiku 5.5, GroundingPI 4B and Muse Wonder.

Google says Gemini 4 is now available to developers in Antigravity and Android CLI, with autonomous self-correction loops, better token efficiency and production-ready Jetpack Compose code out of the box.

A Portuguese-speaking user says that after one more prompt, Gemini 4 Argon delivered a finished 3D piece from scratch, calling it very good at 3D.

Google used Gemini 4 Argon to let its Apollo robot learn cornhole from a single video of a person playing, then land seven consecutive real throws.

A poster says Gemini 4 Argon reaches Astra-level performance on multimodal data annotation, improves significantly over Gemini 3.8, and better identifies actions in egocentric videos.

A reply post suggests that PratikRath15 also test Gemini 4 Argon, part of community calls to benchmark Gemini 4.

A demo shows Gemini 4 Argon creating a browser-only multiplayer capture-the-flag game from a single prompt, with no engine or backend and strangers already playing on opening a tab.