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

The author says the current Gemini 4 Pro checkpoint on Arena has fixed the design taste issues of earlier versions, but after more than 140 tries it is still hard to draw, and he plans to share his unique test cases.

The post rounds up rumors that a leaked Gemini 4 Pro checkpoint reportedly beats Opus 5.5 and rivals GPT-6 Astra and Grok 4.6, while cautioning that benchmarks are not the same as an actual release.

A video of the third Gemini 4 Pro checkpoint on Arena shows a much improved, more polished style over the first version, with more samples and cross-model comparisons promised.

The author ran SVG generation on the new Gemini 4 Pro checkpoint and, after nearly 50 attempts, got a BMW M4 CS side view in Sao Paulo Yellow and Isle of Man Green.

A user says Gemini 4 Pro produced a luxury-style fly image in under 7 minutes from a single prompt, with ruby eyes, metallic wing veins and tiny gears under smoked glass.

A user says they have started to dislike the output formatting of GPT and Claude, and hopes Gemini 4 Pro will be the right choice.

The same prompt was given to GPT-6 Astra and Gemini 4 to each build a dragon flying game with a full flight HUD, wing stamina and altitude tracking, to see which one added more detail.

A new Gemini 4 Pro checkpoint has appeared in the Arena, and the author says Google keeps pushing iterations with visible progress between versions.

Arena quietly added a new Gemini 4 Pro checkpoint with extended thinking; most testers find it worse than the first checkpoint, some call it an upgrade, and the author advises judging multi-run consistency rather than single generations.

xoaanya posts a meme mocking Gemini 4 Pro's long delay, captioned "Gemini 4 Pro benchmarks after a decade of waiting."

VaibhavLab speculates that Gemini 4 Pro is delayed because Google is training internally on the original full architecture on its own TPUs.

HarshithLucky3 says they tried the new Gemini 4 Pro checkpoint 20 times with no output at all and will keep testing.