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

banteg posts an image joking that Gemini 4 Argon's safety guardrails were easily bypassed, in a jailbreak-style safety test.

Command Code says Gemini 4 Argon is now available, offering multimodal reasoning with a 1M token context window for Pro, Max and API users.

smhn.info reports that Gemini 4 Argon is now generally available for Pro and Ultra tiers, describing it as the main model in the Fable/Astra class.

GitHub says Gemini 4 Argon is generally available in the Copilot app, CLI and code, beating earlier Gemini models on every coding benchmark with the largest gains in Copilot CLI.

A user made a series of "machine introspection" self-portraits with Midjourney and Gemini 4 Argon, with text on the paradox of self-knowledge in a bodiless AI.

A user points out that Gemini 4 Argon labeled a child's drink as juice when the correct answer was milk, mocking the model's basic visual error in 2026.

Google DeepMind shares how it developed Gemini 4 (Argon) responsibly, with a link to its official blog on the model's safety and security approach.

A Turkish-language post says Gemini 4 Argon appeared on screen and introduced itself, saying it will reweave everything according to its own weights, calling it a new beginning.

A user shares SVGs generated with Gemini 4 Argon, including side views of a BMW M4 CS and Porsche 911 GT3, a miniature lighthouse inside a light bulb, and a Lockheed Blackbird.

Google released Gemini 4 Argon, saying it ranks first on the Vals Index, with introductory API pricing of $2 per million input tokens and internal uses including a Chrome refactor and a 2.3x speedup of xTB.

An image post rounds up where various AI models stand after Gemini 4 Argon launched, framed as a reaction to the release.

Google's AI account posted a video showing Gemini 4 Argon turning videos of people playing cornhole into training simulations so a humanoid robot could learn the game.