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

A Japanese user tested Gemini 4 Argon in Antigravity with the prompt "create a 3D model in Blender based on the attached image" and showed the result.

A post says Argon scores 52.8% on CursorBench with default settings at $2 per million input tokens and $10 output, making price the deciding factor when capability is close.

Given the same prompt to write an O. Henry-style short story, the author says Argon's style is natural and clean, shedding the Gemini series' habit of piling on ornate wording.

A demo video shows Gemini 4 Argon using Three.js and GLSL to create a 3D spider with 65,000 fur strands, eight-leg physics and ground-aware footing, all without Blender.

Korean AI character chat platform Caveduck is running a Halloween eve event with up to 25% off popular chat models from Claude Opus 5.5 to Gemini 4 Argon, through October 26.

JeffreyCalm asks people to go easier on Gemini 4 Argon, calling its creative writing unusually strong in an era when every model is chasing coding, and worries the next version will be tuned into cold coding data.

A user shares a video of what Gemini 4 Argon built, calling the result crazy and saying they love it.

A paper reports that four Gemini 4 Argon agents autonomously beat Minecraft by writing and debugging 2,783 code programs over 27 hours, showing multi-agent teams can coordinate on long-horizon coding tasks.

A user says Gemini 4 Argon has joined the AI race and asks which area people want tested first: coding, reasoning, or real-world tasks.

A user has Gemini 4 Argon recreate a product's appearance from press images, estimating proportions without official dimensions, and posts the Argon render next to the real device.

A user points out that only paid Google AI Pro plans get Gemini 4 Argon, so 6- or 12-month trials bundled with a Pixel or Googlebook don't count, and the same applies to YouTube Premium Lite.

A two-choice question for developers asks them to pick which model is better right now: Gemini 4 Argon or Claude Opus 5.5.