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

A Russian-language post claims Gemini 4 Argon is fully multimodal, sees, hears and speaks, and after simulation training is ready to deploy on humanoid robots.

Argon is not a fine-tuned Gemini 3 but a new frontier model in the Gemini 4 generation, natively handling text, images, audio and video and aimed at longer coding, research and agentic workflows.

Zhipu's legal model GLM-Legal advertises a 19.2% Harvey LAB score over Claude Fable 5 and GPT-6 Astra, but on the full Vals leaderboard it ranks seventh, behind five Meta Muse models and Gemini 4 Argon.

Gemini 4 Argon has appeared on Arena, where Pro subscribers can try it now.

A video says Gemini 4 Argon is rolling out across Gemini products, aimed at complex prompts and creative tasks, with interest in what developers build.

A heavy GOOG holder who upgraded to Ultra tested Argon: text and diagram output scored 9, but code review produced many hallucinations and fell short of DeepSeek.

A video puts Gemini 4 Argon and Muse Wonder outputs side by side and concludes Argon is clearly stronger.

A promotional post says Gemini 4 Argon has launched for complex multi-step tasks, is available to Google AI Pro subscribers, and links to a third-party order page.

The poster complains that GPT-6 Astra, Opus 5.5, and now Gemini 4 Argon keep launching one after another, each one taking creators' money first.

A user prompted Gemini 4 Argon to generate an SVG portrait of Google AI lead Logan Kilpatrick and shared the result.

A demo shows a single prompt producing an interactive 3D room designer as one HTML file in about two minutes for roughly $1, which the author says could rival Opus 5.5 with a better harness.

Given a single image of a Gothic Victorian carriage, Gemini 4 Argon generated a 174,000-face interactive 3D model in code, with opening doors, hinged windows, wheel physics, suspension, and flickering lanterns.