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

A user shares a screenshot of a new Gemini 4 Pro benchmark, joking that the numbers are too big to display properly.

The same prompt was run on Gemini 4 and GPT-6, with a side-by-side video of the results; the author finds Gemini 4's build very strong.

The author argues Gemini 4 doesn't need to be far smarter, just close to Astra/Fable 5.1 ($10/$50 per 1M tokens) and cheaper, to disrupt frontier model pricing.

A blogger says a suspected Gemini 4 Pro checkpoint is on Arena and collects the strongest outputs seen so far, concluding Google is back.

A poster shares an image with the caption "Gemini 4 Pro?", hinting at another screenshot of output or an interface attributed to Gemini 4 Pro.

A user says Gemini 4 Pro generated a monographic website in 14 minutes, calling it clean and polished and saying earlier Gemini design-taste issues appear solved.

A tester says the Arena model labeled Gemini 3.8 Flash is actually Gemini 4 Pro: fast responses, much better one-shot SVG and 3D generation, and playable small games, though still behind Fable 5.1 and Astra.

LuminaBench notes that the Arena placeholder model "Gemini 3.8 Flash", suspected to route to Gemini 4 Pro, has been renamed "Gemini 3.7 Flash", though it has not verified whether the checkpoint moved with it.

LuminaBench says Gemini 4 Pro generated a PS5 SVG in about 10 minutes, calling it one of the most impressive model outputs so far.

A leaked video shows a game generated by Gemini 4 Pro, with smooth visuals, strong world generation and solid game design, and the poster says Google has something serious in the works.

A user says Gemini 4 Pro produced an interactive 3D model of an Airbus H145 helicopter in about 10 minutes, impressive overall though the bottom UI elements look rough.

Given the same prompt, Gemini 4 Pro and Gemini 3.8 Flash produced pelican-on-a-bicycle SVGs whose gap is called absurd, and a comparison with GPT-6 Astra Pro is promised next.