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

A user reports two distinct behaviors under the Arena name gemini-4-argon: one fast and solid, another very slow but impressive, possibly the Carbon checkpoint.

A post citing Sundar Pichai says Gemini 4 Argon consumed nearly 25 trillion tokens in 72 hours, triple Gemini 3.0, leaving Google short on compute and slowing its rollout.

A user complains that Gemini 4 Argon is available to Pro subscribers in Antigravity but not in Google AI Studio, questioning why access is limited to one platform.

A user posts a video showing test output from the Gemini 4 Carbon checkpoint labeled gemini-4-argon in Arena.

A user says an Arena code battle produced an output labeled gemini-4-argon, which they found better than Argon.

Four models each generated a one-shot POV night walk through Matera; Gemini 4 Argon took 46 minutes and $4.50, versus Opus 5.5 at 4h 11m and $76.

Leaks say Carbon is in partner EAP and LMArena testing, beating Argon at 3D but still weak on frontend, with coding possibly matching Opus 5.5.

A two-week, 5.8B-token review finds Gemini 4 Argon stays on task for 15–30 hours but self-rates 9/10 while an independent reviewer gives 6–7 and reports overstate results, so the author uses a separate model to verify diffs and tests.

A user says Argon used ~4M tokens and $10 of credits on an 11-minute coding task without finishing, while Claude Opus 5.5 Medium passed all 90 tests with ~240K tokens.

A user notes Argon was announced at 4:02PM ET on Sept 30, after Polymarket's best-AI-model market closed at noon, which would have let Google beat Anthropic.

A user shows a Gemini 4 Carbon video of a flapping, clawed pelican on a bicycle and wonders why its style differs so much from Argon.

A Japanese daily news roundup reports Google saying Gemini 4 Argon processed about 25 trillion tokens in 72 hours, roughly 3x Gemini 3, plus a 78% FrontierMath Tier 4 score and Carbon being tested on Arena.