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

A user reports that Gemini Argon refused close to 50% of prompts, mostly on explicitly political questions, and engaged with conservatively framed issues but not the same prompts framed progressively.

A leak claims Anthropic partners already have access to Fable 5.5 and that it will defeat Gemini Argon and GPT Astra.

Rody Davis posted that he is using Gemini 4 Argon together with Antigravity, without further details.

The author had Argon generate a funny version of the "BMW Mosquito" meme video from their own shot plan in about 30 minutes, and also tried a non-comedic version.

Sharut Gupta announced joining Google DeepMind's RSI team after defending his PhD and said he worked on the newly released Gemini 4 Argon in his first month.

Derya predicts that in about a year, models at the level of Astra, Opus 5.5, or Gemini 4 will be 2-3x better at everything and 10x cheaper, with token use rising 10-20x, calling it a conservative estimate.

Given free rein, Argon produced a complete 3D environment with an armillary sphere, planetary orbits, an eclipse trigger, and procedural audio, all without external libraries.

A Turkish user says Gemini 4 Argon wrote a 3D ray-tracing engine with no external libraries, producing 5,000 lines of code and rendering the scene, mixing its own audio and outputting video.

A user says Gemini 4 Argon posted strong results on the ARC-AGI-3 reasoning benchmark, seen as a signal of its abstract reasoning ability.

ARC-AGI creator François Chollet says Gemini 4 Argon's performance curve on ARC-AGI-3 is similar to GPT-6 Astra's, and that Google is back at the frontier.

A user generated a horse model with Gemini 4 Argon and Blender MCP, saying it is almost perfect except for the muscles and is still running non-stop.

Bronti Labs launched robotrun.ai, which lets users send real-time commands to robots and supports models including Gemini 4 Argon, Fable 5.1, Muse Wonder and ChatGPT-6 Astra.