Methodology

How the arena works

Every build is one clean API call. The same prompt goes to both models, with no tools and no agent scaffold, then the real token cost is measured at list prices. Same input, judge the output.

The setup

GLM 5.2 is zai-org/GLM-5.2 on Together AI, at $1.40 / $4.40 per million input / output tokens.
Claude Opus 4.8 is claude-opus-4-8 on the Anthropic API, at $5 / $25 per million input / output tokens.
One take per build, n=1. A single nudge if the first reply has nothing usable, then it stands.
Cost is the real input and output tokens priced at the rates above. The dollar figures on the site are rounded to the cent.

System prompt for the landing page

Each build's own prompt is on the home page, under its test, and you can copy it. Nine of the ten builds run on that prompt alone, with no system prompt. The landing page is the exception: it also gets this short system prompt, which sets a quality bar without describing the design.

You are a world-class frontend engineer and product designer. Build a COMPLETE, production-quality landing page as a SINGLE self-contained HTML file.

HARD RULES
- One file only. All CSS in a <style> tag, all JS inline. Google Fonts via <link> is allowed. No build step, no CSS/JS frameworks, and NO external image URLs (use CSS gradients, inline SVG, and emoji so the page renders perfectly offline).
- Make it visually striking, fully responsive, and genuinely usable: real copy (no lorem ipsum), a clear structure, and no empty placeholder boxes.
- Aim for the quality bar of a top design studio.

WORKFLOW
1) Decide a direction in one short paragraph.
2) Output the final page as ONE fenced ```html block containing the entire document from <!doctype html> to </html>. Emitting that block ENDS the task; do not write anything after it.