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

A developer asked Gemini 4 Argon to build Ancient Egypt in ThreeJS and got a full, surprisingly realistic runnable 3D world, shown in a video demo.

A leak claims three OpenAI codenames, aurora red/green/blue alpha, which the author guesses are tiers of one generation under blind testing and plans to compare against Gemini 4 Argon.

Google's blog says one tabletop photo becomes a simulation twin running 4,500+ randomized episodes, lifting real-robot success from 56% to 75% and cutting time from 35s to 24s.

A short reply praising Gemini 4 Argon, with no specific test scenario or numbers given.

A user asks whether anyone has tried Gemini 4 Argon for coding and how it compares to Opus 5.5, with no answers in the post itself.

A video post compares Gemini 4 Argon with Opus 5.5, GPT-6 Astra and Sonnet 5.5, with the author saying they've reached a conclusion.

Self-reported figures say Gemini 4 Argon refactored a 6M-line folder in Chromium's 40M-line codebase in weeks instead of months, found 16+ timeout and race-condition bugs across 700+ internal labs, and saved 955 engineering hours.

A Korean web-novel author says cases of AI-automated web fiction and worldbuilding are appearing after Gemini 4 Argon's release, and calls for industry AI usage guidelines.

A user saw an Antigravity banner claiming Gemini 4 Argon is available to Pro subscribers but still can't access the model, and asks whether other Gemini Pro users can.

Google's Antigravity team confirmed it is dropping the standalone VS Code fork, so even paid users lose Argon there; users must switch to vanilla VS Code with the official extension or the Antigravity 2.0 app and agy CLI.

A video roundup of this week's AI releases, with Gemini 4 Argon listed alongside Krea 3, Kling 4, and the Muse lineup.

A heavy Gemini 4 Argon user says the model answers ambiguous prompts with confident, unchecked claims, raising concerns about sycophancy and hallucination.