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

A poster argues that Gemini 4 Pro and Astra produce nearly identical results because the benchmark used is saturated at the top and can no longer separate models.

A user reading LuminaBench results says Gemini 4 and GPT-6 Astra are extremely close, but notes that Gemini drew the logo correctly.

On the same SVG prompt, Gemini 4 Pro and GPT-6 Astra both come out near perfect, though Astra's logo is slightly off, and the gap to other models is large.

A joking post says Gemini is no longer chosen by vibe coders and that only the release of Gemini 4 Pro can save Google CEO Sundar Pichai.

A user reran the same prompt in Arena battle mode and got identical outputs labeled muse spark 1.3 and kimi k3, raising suspicion that a model is simply being relabeled as Gemini.

Replying to LuminaBench, a user says Gemini has long been decent at SVGs but Gemini 4's output is clearly a step cleaner.

The author says "gemini 3.7 flash" in Arena, which was routing to Gemini 4 Pro, took about 20 minutes to draw a complex Xbox controller SVG, but the same result could not be reproduced the next day.

A model listed on Arena as gemini-3.7-flash produced highly polished scenes and UI, and the entry was later removed, with the output said to be a Gemini 4 Pro test.

Google product lead Tulsee Doshi says her team is excited about the unreleased Gemini 4 and building internal hype while balancing the chaos and possible burnout of moving fast.

A user asked Gemini 4 Pro to build a mechanical butterfly in three.js, and the reposted result is called surprising again, though the post gives no code or render details.

A video shows Gemini 4 Pro's three.js generation, which the author rates on par with Fable 5.2 and says makes them more eager for Gemini 4.

A short post argues that Gemini 4 Pro may reverse its earlier reception, without citing any tests or data.