Two new AI models. One prompt. One planet.
OpenAI released GPT-6 Sol and GPT-6 Luna today. We ran them twice: once with only the brief, then a fair rematch with exactly the notes Opus had. Then we reran Opus with the brief only. We gave each one the same one-line brief we gave Claude Opus 5.5 earlier tonight. Each wrote one Three.js page on its own, in one go. We didn't nudge, retry or fix anything. What you play below is exactly what each model wrote.
output a ThreeJS 3D cartoony/comics-like planet, with vibrant life on it: clouds, mountains, city buildings, cars and plane romaing it. It should be possible to orbit around and zoom and see the details. No harness is to be used
Round 2: the fair rematch
Opus had a head start in round 1: before building, it had read our project instructions and notes, including what we liked in an earlier planet (plane quality, two-lane roads, a round globe, polished clouds). So we reran Sol and Luna with exactly what Opus had read before its first line of code, word for word, and the same request. Nothing more, and still one prompt with no retries.
GPT-6 Sol, round 2
Fullscreen ↗GPT-6 Luna, round 2
Fullscreen ↗| Model | Build time | Build cost | Looked at its own page? | Page size |
|---|---|---|---|---|
| GPT-6 Sol, round 2 | 7.1 min | $0.70 | yes, desktop and phone | 24 KB |
| GPT-6 Luna, round 2 | 6.0 min | $0.04 | yes, and fixed a broken import it found | 20 KB |
| Claude Opus 5.5 | 14.0 min | $2.54 | yes, three times | 48 KB |
Build time runs from the start to the last change to the page, the same window used for Opus. Cost is at each model's published API price.
GPT-6 Sol, round 2
What changed with the same notes
- Sol jumped the most. Dense forests, four cities with ring roads, three flight paths, a pause and reset button. It checked its page on desktop and phone before finishing. It is now close to the Opus page in richness, at about a quarter of the cost and half the time.
- Luna got better too. Three sky cities, visible air routes, and it found and fixed a broken module import by looking at its own render. A visible square frame around the globe still spoils it.
- Opus still wins on the details up close. Its camera tilts toward the horizon as you zoom, so you end up flying over roads and hills. Both GPT-6 pages zoom straight down into flat ground.
Round 3: Opus with the brief only
To close the loop, we reran Claude Opus 5.5 the way Sol and Luna ran in round 1: a clean install with no project instructions or notes, the same one-line brief, one prompt, same thinking effort.
Fullscreen ↗
| Brief only (clean install) | Build time | Cost | Looked at its own page? | Page size |
|---|---|---|---|---|
| Claude Opus 5.5 | 14.1 min | $2.69 | no, syntax check only | 50 KB |
| GPT-6 Sol (round 1) | 4.4 min | $0.24 | tried, no working browser | 22 KB |
| GPT-6 Luna (round 1) | 2.0 min | $0.005 | no | 15 KB |
- Opus is the only one that really went comic without our notes: ink outlines, halftone dots, comic-book buttons, continents with mountains and a round city, windmills, and a zoom that tilts to the horizon.
- It spent its time thinking, not looking. Its first step was one 13-minute burst: about three quarters of its 93k output tokens were thinking, then it wrote the whole page in one go. It never opened the page in a browser.
- Price gap: about 11x Sol and 500x Luna for the brief-only page.
Round 1: the brief only
The first run: each model got only the one-line brief. Drag to spin, scroll or pinch to zoom.
GPT-6 Sol
Fullscreen ↗GPT-6 Luna
Fullscreen ↗Time and cost
Cost is what the tokens each model used would cost at OpenAI's published API prices. Sol: $2 per million input tokens, $0.20 cached, $10 output. Luna: $0.10, $0.01 and $0.50.
| Model | Time | Input tokens (cached) | Output tokens (of which thinking) | Commands run | Page size | Cost |
|---|---|---|---|---|---|---|
| GPT-6 Sol | 4.4 min | 240,583 (206,976) | 13,397 (4,641) | 10 | 22 KB | $0.24 |
| GPT-6 Luna | 2.0 min | 66,958 (56,320) | 6,140 (788) | 3 | 15 KB | $0.005 |
| Claude Opus 5.5, same brief, earlier tonight | 14.0 min | 56.6k | 48 KB | $2.54 |
Stills
Every page was shot the same way: at rest, after a drag to orbit, and after a few scroll steps of zoom. There were no errors on any page.
GPT-6 Sol
Our take
| Good | Weak | |
|---|---|---|
| GPT-6 Sol | Looks like a finished product page. Clean low-poly style, a round globe, a walled city, a plane with its flight path, and working pause and orbit buttons. | Not much life yet: one or two towns, few cars, no roads. Zooming in shows flat ground rather than details. Not very comic-like. |
| GPT-6 Luna | Works first time, moves, looks bright. It cost half a cent. | Most of the land and all the city sit on one edge of the globe. Zooming in shows clouds and sea. Cities are stacks of boxes. |
| Claude Opus 5.5 | Still the richest page: ring roads, lots of cars and planes, comic ink lines, a camera that tilts as you zoom so details stay visible. | About 10x Sol's cost and 3x its time. |
In short: Sol gets you most of the way to the Opus page for a tenth of the price. Luna makes a working toy for almost nothing. Note that Opus had a head start (see below).
How it was run
- One prompt, no follow-ups. The brief above, word for word, plus one line saying where to save the file.
- A clean install. Each model ran in OpenAI's Codex command-line tool with no project instructions, notes or add-ons. The only prompt was Codex's own built-in one. Thinking effort was set to high, the same as the Opus run.
- Same computer as Opus. A full shell and internet access (three.js loads from a CDN). The two models ran at the same time.
What this does not show
- Round 1 only: Opus had a head start (round 2 removes it). The Opus run was not blind. It ran inside our own coding setup, with our project instructions, notes and tools already loaded. Before building, it read our notes on what we liked in an earlier planet: plane quality, two-lane roads, a round globe, polished clouds. Sol and Luna saw only the brief. So Sol vs Luna is a fair fight; Opus vs GPT-6 is not yet.
- In round 1, Sol never saw its own page. Partly our fault: the clean setup we used also hid the working screenshot tool, which Opus had. Round 2 fixed that, and both models checked their pages.
- One run each. A second run of the same prompt could come out better or worse.
- No real-GPU frame rate measured. The stills are software-rendered. Not tested on a phone.
- Needs internet: three.js loads from jsDelivr.
Made by OhMyUnicorn Forge: describe an app in plain words, get a live URL.