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

The author asks Argon to design a fidget spinner printable in under an hour, a task he says Opus 5.5, Astra and Grok 4.7 all failed, and tests it on a Bambu Lab P1S.

Each model got 25 minutes to build a fighter for a Super Smash Bros-style fight, and Argon won all three matches without a loss.

The author says Argon is far faster than other frontier models, cheaper than Opus 5.5 on some tasks, strong at writing and 3D work but weak at front end, and some find it as indecisive as GPT models.

A user issues a taunting challenge asking people to name any AI model better than Gemini 4 Argon, backing the claim with an image.

A developer who received a Devin subscription planned to test Gemini 4 Argon with it but says the model is so boring and mediocre that they have no enthusiasm to test it even for free.

A user says the circulating Argon gummy bear video is fake, reruns the same prompt, and finds Gemini 4 Argon much better than the fake though still behind Opus.

A Chinese user finds Gemini 4 Argon strong enough to produce Opus 5.5-style video effects but says Google's Antigravity platform is weak and has made little progress, dragging down the model.

The author showed Argon the DeepMind pagoda demo and asked for a walkable interactive version; over several iterations it added controls, a koi pond and bell-ringing on its own, taking about 4 hours.

The author asked Gemini 4 Argon to propose a unifying theory of biology by 2040 linking immunology and neuroscience through first principles and math, then had it turn the theory into a poster and a simpler explanation.

Hermes now offers Gemini 4 Argon with long-decode support past the native 64K output cap, plus phone control, a mobile adapter and cross-session handoffs.

The poster says they asked Gemini 4 Argon to fix a bug and, after about 30 minutes, the result still carried Gemini's familiar quirks.

A Turkish-language test of Gemini 4 Argon across 13 tasks found it smart but inconsistent and hard to steer, losing a 3D world test to Opus and hitting usage limits faster than Claude.