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

LisanBench added multi-reasoning-effort tests: Gemini 4 Argon low and medium emit no reasoning tokens yet score high, with low taking #1 on the non-thinking board despite being a dirty looper, and Argon high ranking second without the edit-distance rule.

HarshithLucky3 shows an early-stage attempt at driving Blender via MCP with Gemini 4 Argon, an initial foray into 3D modeling.

A tester says Kimi K3.1 Pro is a big quality jump with fewer hallucinations and lands around Gemini 4 Argon, with each model winning different tasks.

A user measured that the $100 Google AI Ultra plan maps to roughly $140 of Gemini 4 Argon usage, mocking the value versus Codex and Claude quotas and calling Gemini 3.8 Flash a freebie.

A user shares a Google Antigravity usage panel for Gemini 4 Argon High showing 67% of the weekly limit still remaining with a reset in 7 minutes.

The same water-pump prompt went to Claude Opus 5.5, Gemini 4 Argon, GPT-6 Astra and Grok 4.7, yielding four very different mechanical designs with no scoring, leaving readers to judge which actually works.

A user posts their first Gemini 4 Carbon result from Arena and says more tests will follow if the checkpoint keeps working.

A discussion post ranks Gemini 4 first, GPT-6 Astra second, Grok 4.7 third and Claude Opus 5.5 fourth, asking which one actually wins.

A demo pairs Gemini 4 Argon with Blender MCP to convert a single Nano Banana-generated Gothic cathedral image into a 3D scene, showing AI directly driving modeling tools.

Author ran six Japanese writing tasks across Gemini 3.8 thinking mode, 3.6 Flash and 4 Argon; only 3.8 hit exactly 200 characters, while Argon excelled at plain-language explanations for 70-year-olds.

A video argues GPT-6 and Gemini 4 Argon didn't simplify building; they made it harder for founders to decide which model fits which task.

A short video jokes that after Kimi K3.1 Pro, Gemini 4 Argon and Claude Haiku 5.5, the bug is still alive.