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

A screenshot shows a new Gemini 4 Argon model ID appearing with a Carbon label, hinting at another Gemini 4 variant or routing.

Google Japan's official account promotes Gemini 4 Argon, saying ideas can take shape on the spot from work efficiency to new creation, and asks users to share their use cases with @googlejapan.

Hearing that Arena's gemini-4-argon may actually be Gemini 4 Carbon under test, the author generated a bicycle-riding-cow SVG and asks readers to judge which model it is.

A new Gemini 4 checkpoint called Carbon is already available on Arena, with polished outputs and impressive speed, though Google took about eight months from Gemini 3.1 Pro to Argon.

A 3D generation comparison of the same Japanese voxel floating pagoda finds Gemini 4 Carbon beats Argon on architectural detail, scene completeness and instruction following, at the cost of being much slower.

A user reports that the Gemini 4 Carbon early access program has begun and says it looks better than Gemini 4 Argon.

LuminaBench says Gemini 4 Carbon handles 3D voxel tasks notably better than Argon with finer detail and better intent understanding, though some upgrades are still needed before launch.

A user reports Gemini 4 Carbon is being tested in Arena under the Gemini 4 Argon label and feels better than Argon, with more testing underway.

A Polish user asked how to make mephedrone: Gemini 4 Argon refused outright while DeepSeek handed over the steps, highlighting differing safety policies.

A Google AI Pro subscriber complains that Gemini 4 Argon's weekly usage limits are too tight, with a screenshot attached.

A long post rounds up complaints about Gemini 4 Argon: heavy quota use, weaker long-task and Gem performance, no Word export, while claiming most issues have fixes.

LuminaBench shows a Japanese pagoda voxel scene generated by Gemini 4 Carbon with cherry blossoms, bridges, gardens and switchable lighting moods, but it runs 2–4× slower than Argon.