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

A demo video shows Gemini 4 Argon writing a working 3D model of a futuristic combat helicopter in Three.js from scratch, with no pre-made assets.

Japanese outlet ITmedia covers the general release of Gemini 4 Argon and outlines the plans and pricing where it is available.

Parallel Search added nine new models to its capability leaderboard, with Gemini Argon 4 at #5 and Sonnet 5.5 at #2.

A user woke up to find Gemini 4 Argon available in the app on their Google AI Pro plan and plans to check whether it also works in the TUI.

A user asked Gemini 4 Argon to build an exact 3D replica of the 1966 USS Enterprise from Python code in Autodesk Fusion and Blender, then render a video with a Python-synthesized Star Trek theme.

Wes Roth notes Argon ranks fifth in Agent Arena but second in steerability and third in praise versus complaints, and says the uncertainty ranges overlap with competitors, so the focus should be on whether it cuts back-and-forth in real workflows.

davis7 posts an image showing their full experience using Gemini 4 Argon, reportedly Gemini 4 Pro.

A user says Gemini 4 Argon is now usable in Droid, marking the model's entry into third-party coding agent toolchains.

Google released Gemini 4 Argon, with Pro users able to use it only in the Gemini app while Ultra users can call it directly in Antigravity.

A short hands-on post about trying gemini-4-argon, with images showing the actual output.

Pydantic AI now supports Gemini 4 Argon long decode: one request stops at 32K output tokens (64K streamed), and it follows Google's continuation token to return one combined response.

The author says they tried Gemini 4 Argon for image generation and, after several rounds of back-and-forth, found the rate limits looser than Opus used to be.