DEV Community

Diwakar Ray Yadav
Diwakar Ray Yadav

Posted on

I tested Gemini 4 Neon on 4 creative-coding prompts. Here's the raw output

Google's Gemini 4 Neon dropped this week, ranked #3 in the world on the Artificial Analysis index. I gave it four of my hardest creative-coding prompts in AI Studio (Neon High), one shot each, with no edits to the output.

Test 1: Kathmandu rooftop at night

The prompt: a night-time Kathmandu rooftop scene with prayer flags in the wind, a glowing Boudhanath stupa, lanterns drifting up, bloom and fog, made to look like a film still.

Result: 4,783 lines in one HTML file, three.js only, everything procedural. Watch the recording

Test 2: a 60-second reel, fully coded

"The Life of a Prompt": a sentence typed on a phone flies to a data centre, gets split into tokens and comes back as an answer. Three colours, hard cuts, a 120 BPM beat made with Web Audio.

Result: 2,655 lines, GSAP as the only dependency. Watch the recording

Test 3: a walkable Himalayan village

Stone houses with smoking chimneys, a river and a bridge, prayer flags, a chorten, a prayer-wheel wall, and WASD plus mouse-look controls.

Result: 2,291 lines, three.js only, no assets. Watch the recording

Test 4: the pelican on a bike, neon edition

One SVG with no JS, images or fonts: a pelican riding through a neon city at night.

Result: 957 lines and 45 SMIL animations. The wheels spin, the sign flickers and the wet asphalt reflects. Watch the recording

Verdict

4 prompts, 10,686 lines of code, zero assets. Neon is strongest on mood: its light, fog and motion design look art-directed rather than generated. Its 3D worlds look great but get a bit low poly up close. For creative coding, it's a real contender.

Every prompt, raw output and recording is open source, so you can run the same tests on any model: https://github.com/Diwak4r/model-reviews

I test frontier models so you don't have to. Follow me on X at @Norwakar.

Top comments (0)