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

Gemini 4 Carbon is reportedly in teamfood internal testing on Antigravity and already on Arena, with early performance said to be comparable to Opus 5.5 and a possible launch in the first week of next month.

A user compares Argon, Opus 5.5 and GPT-6 Astra: Argon explains complex code clearly and holds long context well, but its constant flattery makes it unusable for learning.

The author argues Gemini 4 Argon is strikingly different: while other models merely think and reason more when effort goes from medium to high, Argon behaves in a way that feels almost AGI-like.

A user had Gemini 4 Argon generate a 3D Infinity Castle from Demon Slayer with floating rooms and endless stairs, and asks when Gemini 4 Carbon ships.

The same prompt asks four models to animate a two-rotor lobe pump whose rotors must mesh exactly; the author scores nothing and asks viewers which one actually pumps.

A Japanese post says Google released Gemini 4 Argon after ten months, first to vetted cyber-defense users, raising the single-response cap from 64K to 1M tokens at $2 input and $10 output per million.

A user feels Gemini 4 Argon's usage limits are nerfed compared with Opus 5.5, unsure whether limits changed or they are simply burning more tokens.

One prompt, four models build a trebuchet, comparing counterweight, sling and release timing to see which physics holds up.

The author suggests letting Gemini 4 Argon and 3.8 Flash handle big-picture planning, then sending results to bundled Opus/Sonnet 5.5 or external models for review to offset hallucinations and sloppiness.

A rumor says Gemini 4 Carbon is in testing with gains over Argon in 3D generation, spatial reasoning and bulk throughput, coding compared to Opus 5.5, though slower on some tasks.

An image post shows a plane rendered by Gemini 4 Argon, with the poster calling its image generation unreal.

Thailand's AiPASS platform launched a Gemini Enterprise program granting 500,000 Thais free access to Gemini 4 Argon for 10 months, with daily quota resets, image/video generation, Deep Research and 50+ agents.