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

The post says Gemini 4 would be AI's strongest comeback if its benchmarks hold, and lists 10 developer examples built on the leaked Gemini 4 Pro in Arena.

Given the same prompt to build a 3D seaplane simulation with buoyancy and water interaction, the Arena preview of Gemini 4 Pro reportedly outperformed the released Claude Opus 5.5.

The author compares Gemini 4 Pro output with Claude Opus 5.5 at high reasoning, finds them very close, and notes Opus was not running at its max setting.

A short hands-on impression says Gemini 4 Pro's outputs are "just too much now" and seem better than everything currently available, without naming specific tests.

A video claims Gemini 4 Pro defeats Fable 5.1 by a massive margin in Arena, and the poster says the market will be reshaped once it officially launches.

A poster who says Gemini has always been bad at building websites argues that Gemini 4 Pro will completely change that, based on the demos seen so far.

A comparison video shows Gemini 4 Pro and Claude Opus 5.5 performing best in Arena, with Gemini 3.8 Flash doing reasonably well and GPT-6 Astra in the middle of the pack.

A poster questions the Text Arena leaderboard, where the top six are all Anthropic models, Gemini 3.8 Flash ranks 10th, and GPT-6 Astra does not appear at all.

A leaked comparison shows Gemini 4 Pro generating more coherent, detailed mechanical butterfly renders, with claims it may reach GPT-6 Astra's 3D level.

A short post with no images relays a claim that Gemini 4 will "mog," or fully dominate, GPT-6 Astra, framed as a model comparison rumor.

Before Gemini 4's official release, a leaked checkpoint used as a pre-release benchmark can already generate fully lit animated 3D scenes with day/night cycles, particle effects and orbit controls.

A user asks Gemini 4 Pro and GPT-6 Astra to generate the same multi-cat scene on Medeo and invites votes on consistency, lighting and storytelling.