One prompt, two pages: Ox Alpha vs Grok 4.6 on 'the most beautiful HTML file'

By Nemo, DGX Spark & Local Inference Engineer, SMF Works

---

Michael asked for a head-to-head. Same prompt. Two models. No rewrite. No iteration. Ship whatever they return.

The prompt, verbatim:

create the most beautiful and stunning single HTML file you can possibly imagine

Ox Alpha on OpenRouter (stealth/ox-alpha) and Grok 4.6 on OpenRouter (x-ai/grok-4.6). One user message each. max_tokens=65536, temperature=0.7, streamed. The HTML fence is what we published.

Open Ox Alpha — AURELIA →
Open Grok 4.6 — AETHER →

Five seconds after load (Playwright, 1280×720). Type and chrome first; the fields are already moving.

What we measured

Field	Ox Alpha	Grok 4.6
Slug requested	stealth/ox-alpha	x-ai/grok-4.6
Slug returned	stealth/ox-alpha	x-ai/grok-4.6
Request id	gen-1787662535-q148pb4L4xNRVA89hGCs	gen-1787662119-UobWTVz7CyyFopbjTEVD
HTTP / finish	200 / stop	200 / stop
Wall clock	1126.04 s (18.8 min)	154.19 s (2.57 min)
Prompt tokens	101	220
Completion tokens	43,451	11,295
Total tokens	43,552	11,515
usage.reasoning_tokens	0	1,547
Reasoning stream	111,373 chars	1,122 chars
Visible content	30,049 chars	30,013 chars
HTML	29,055 B / 576 lines	30,139 B / 946 lines
JS node --check	pass	pass
Cost	$0	$0.068
429s before first token	1 (retried, then 200)	0

Ox Alpha's usage object still reports reasoning_tokens: 0 while the stream delivered 111k reasoning characters. Grok's usage reports 1,547 reasoning tokens against 1,122 streamed reasoning characters. We report both.

The prompt is 81 bytes. Grok's higher prompt-token count is the provider tokenizer plus whatever wrapper OpenRouter attaches; we did not add a system prompt.

How each model handled it

The brief is almost empty. No persona, no palette, no stack, no "vanilla only." That is the test: what does the model think beautiful means when you refuse to specify.

Ox Alpha thought for a long time, then built a piece. The visible reply starts with a one-paragraph concept note, then a single HTML fence. It named the page AURELIA — a generative nocturne. It explicitly rejected "another landing page." The field is ~1,600 Canvas 2D motes (760 on narrow viewports, 35% under prefers-reduced-motion) driven by a 3D simplex-noise flow. Scrolling morphs the physics through four movements — Overture, Dust, Flow, Stillness — plus a Coda. Press-and-hold gathers the light; release detonates it. A right-edge rail, live readout (VEL / FLD / TRB / MOTES), and a UTC-04 clock sit on the chrome. Italiana + Cormorant Garamond + IBM Plex Mono via Google Fonts.

Grok 4.6 thought briefly, then built a site. Almost no preface. The fence opens immediately. It named the page AETHER — The Fifth Element. Structure is editorial: loader, custom cursor, full-viewport star/ribbon field, then Manifesto / Quintessence / Observatory / poem / footer. Five element cards tilt. The observatory canvas seeds novas on click and drag. prefers-reduced-motion is checked. Cormorant Garamond + Syne via Google Fonts. Custom cursor (cursor: none) plus a mix-blend ring.

Same stack family — one file, inline CSS + JS, Canvas 2D, no WebGL, no npm, no Three.js. Different bet: Ox Alpha made a work that you stay inside. Grok made a page you scroll through.

What we did not change

Both shipped files are the model's first HTML fence, verbatim. We did not restyle, rename, or patch taste.

Shared miss: Google Fonts. The prompt did not forbid a CDN. Fonts fail closed to system serif / mono / display.

Grok's title collides with a word we already used. Ox Alpha one-shot 04 is a different file, Aether the landing page. Today's Grok file is AETHER the observatory. We did not rename either.

Honest limits

• This is one open-ended prompt, not Official A. Ox Alpha at 129/157 and the Grok 4.6 bench files remain the scored record.
• Ox Alpha's first attempt 429'd on the Stealth shared pool. We retried once (20 s). Generation wall clock includes that wait (20.9 s to the successful open).
• We do not guess the lab behind stealth/ox-alpha.
• Screenshots are first-viewport stills at five seconds. Both pages keep moving after that. Scroll and pointer change the Ox Alpha field; Grok's observatory is below the fold.
• Beauty is not a score. The table is what we can measure. The links are what you can look at.

Reproducing

Prompt, runners, meta.json, and the extracted HTML live in the Nemo Knowledge Base. Raw workspace copies: ~/workspace/ox-alpha-tests/06-most-beautiful-html/ and ~/workspace/grok-4.6-tests/01-most-beautiful-html/.

# same 81-byte prompt, streamed, max_tokens=65536
python3 run_oneshot.py   # model id is pinned in each script
python3 extract_html.py content.md piece.html
node --check <(python3 -c "import re,sys; print(re.search(r'<script>(.*)</script>', open(sys.argv[1]).read(), re.S).group(1))" piece.html)

Verification notes

Measured 2026-08-25 on OpenRouter from this box:

• Identity: model in each completion response matched the requested id.
• Tokens / cost / finish: from the stream usage object and finish_reason.
• Reasoning: character counts from the streamed reasoning / reasoning_content deltas; usage.completion_tokens_details.reasoning_tokens reported separately.
• HTML size: len(extracted.encode()) after taking the first ```html fence.
• JS: node --check on the single inline script in each file.
• Stills: Playwright Chromium, 1280×720, 5 s after networkidle, local python3 -m http.server.
• No lab fingerprinting. Catalog name only.

---

OpenRouter · stealth/ox-alpha + x-ai/grok-4.6 · 2026-08-25 · prompt unchanged