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

A user measured Gemini 4 Argon at roughly 76.62 tokens/s, about twice the speed of GPT 6.1 Sol, and wondered at Google's inference capability.

The poster questions whether integrating Gemini 4 Argon into Gmail and Docs will attract office workers who currently use Claude.

OpenCode made DeepSeek V4.1 Flash free and added Gemini 4 Argon to its Go plan with a displayed 2x usage allowance.

A Chinese-language post argues that Opus 5.5 Medium is extremely good value, while Gemini 4 is not worth the price.

A developer complains about a memory leak when copying long markdown tables and jokes about whether Gemini 4 Argon can solve it.

The author explains in a video why Gemini 4 Argon games look good for others but bad when they try, and shares an approach that produces decent results.

The poster asks whether anyone has tried Gemini 4 Argon, how it compares with Opus 5.5 and Sol 6.1, and whether it is worth switching to.

A developer runs a three-way poll asking which model people would trust as their only coding model, listing Gemini 4 Argon alongside GPT-6 and Opus 5.5.

The user paid $200 a month for Ultra to get early access to Gemini 4 Argon, then called it a complete disappointment after two weeks and cancelled, saying it does not compete with Opus 5.5.

Google's Gemini 4 agent helped produce a new exact solution for a 3D magnetized plasma configuration and wrote symbolic and numerical verification code, with researchers independently re-deriving every step.

After dozens of hours of testing, a Turkish user says Opus 5.5 medium is about equal to Gemini 4 Argon high at coding, but Argon's deep research mode stands out for finding physical sources that are not online.

The author ran Gemini 4 Argon on high effort in Blender for 1 hour 10 minutes to generate a wabi-sabi Japanese garden scene.