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

A user says the gemini-4-argon entry in Arena now runs a new Carbon checkpoint and is starting to test it.

LuminaBench says Gemini 4 Carbon has been in Arena since yesterday, showing voxel pagoda outputs that beat Argon but run 2–4x slower depending on the task.

The post claims Gemini 4 Carbon has reportedly appeared on LMArena and been given early access to partners, hinting a public release may be near.

Leaks and early tests say Carbon was tried on Google's internal Jetski coding platform, described as feeling like Opus 5.5, with clearly better 3D and less laziness than Argon.

Koray Kavukcuoglu says Gemini 4 Argon handled nearly 25 trillion tokens in its first 72 hours, three times Gemini 3's volume.

A user reports that Claude treats Gemini 4 Argon, not GPT-6 Astra, as its biggest competitor in overall rankings.

A leak says Gemini 4 Carbon is in partner EAP and on LMArena, beats Argon at 3D but still weak at frontend, with release possibly late October or early November.

Rumor says Gemini 4 Carbon is in early partner testing with stronger 3D than Argon, possibly launching this month, though public access may take much longer.

A user complains that Google AI Pro's Gemini 4 Argon quota is just 60M+ tokens, while the same price buys 800M+ tokens of 6.1 Sol on ChatGPT Plus or 3B tokens of Opus 5.5 on Claude Pro.

A user says they haven't finished testing Gemini 4 Argon and already hears the Carbon codename model is better at 3D, joking Google is moving too fast.

The post says Gemini 4 Carbon is already available to enterprise users and reportedly beats Argon in 3D modeling, though Argon is already strong.

The poster guesses the newly surfaced Carbon is a smaller Flash- or Pro-class model with Argon as Google's Astra/Fable equivalent, and credits Sergey Brin for the faster cadence.