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

Gemini 4 Argon is now available in Google AI Studio for both Pro trial and paid plans, with usage limits, and the author asks what people will build first.

A research team ran Gemini 4 as multiple agents on Google Antigravity and, within hours, proposed two new analytic solution families for a plasma confinement problem; researchers independently verified the derivations.

Gemini 4 Argon reads and rewrites a dense paper into a plain narrative, Nano Banana illustrates each section, and Veo animates the images.

Pro subscribers complain that the Argon model exhausts their usage after just two or three prompts, questioning what they pay for.

Users report Google Antigravity pushed a global weekend quota reset for Gemini 4 Argon and Gemini 3.8 Flash pools, restoring weekly limits to 100%.

A post claims four Gemini 4 Argon agents completed Minecraft in 27 hours with zero human help, without citing a source.

A Japanese user had Gemini 4 Argon build a shared-timeline slider tool to compare before/after mixing and mastering results.

A user tests a cheap Google AI subscription: Argon works in the Gemini app but not Antigravity, which offers Gemini 3.8 Flash and Opus 4.6 instead.

A user posts an image showing that Google reset their Gemini 4 Argon usage in Antigravity, expressing confusion at the unexpected reset.

DeepMind let four Gemini 4 Argon coding agents play Minecraft from scratch with no assigned roles; after 27 hours, 605 deaths and ~$520 in API cost they killed the Ender Dragon.

A tally shows Gemini model lifespans shrinking from a year to weeks, with 3.8 Flash lasting 13 days, Gemini 4 Argon three days old, and a Carbon model already in internal testing.

A post lists Gemini 4 Argon, Gemini 4 Carbon and Claude Opus 5.5 together, joking that tracking AI model releases has become a full-time job.