CYPHIR Benchmark
60% of AI-generated code
is human overhead.
CYPHIR is the AI-native authoring format.
Same web output. Fewer tokens. Better code.
Your token limit just got 2.5x longer.
The overhead exists in all web code. AI is the first author that doesn't need it — so we built a format that doesn't produce it.
About CYPHIR
CYPHIR is an AI-native authoring format. It replaces the verbose, human-readable code that AI models currently generate with a compact intermediate representation that compiles to standards-compliant web output.
The thesis is simple: 60–80% of AI-generated web code exists solely for human readability. Eliminate that overhead, and you reduce token costs, latency, and error surface — while the browser receives identical output.
A Layer-by-Layer Diagnostic
Every layer of the web stack carries human-readable overhead that AI doesn't need. The warm-toned segments represent tokens that exist purely for human comprehension.
Same output.
A fraction of the tokens.
Seven UI components. Two authoring approaches. Identical visual output. The token counts speak for themselves.
Never hit the token wall mid-build again.
low — A single content card with image, title, description, tag/badge, and CTA button. Includes hover states.
<div class="max-w-sm rounded-xl overflow-hidden shadow-lg bg-slate-800 hover:shadow-2xl transition-shadow duration-300">
<img class="w-full h-48 object-cover" src="/image.jpg" alt="AI-generated abstract visualization">
<div class="p-6">
<span class="inline-block bg-indigo-500/20 text-indigo-400 text-xs font-semibold px-3 py-1 rounded-full mb-3">
Research
</span>
<h3 class="text-xl font-bold text-slate-50 mb-2">The Human Overhead Thesis</h3>
<p class="text-slate-400 text-sm leading-relaxed mb-4">
60–80% of AI-generated web code exists purely for human readability. We measured it.
</p>
<button class="w-full bg-indigo-600 hover:bg-indigo-500 text-white font-medium py-2.5 px-4 rounded-lg transition-colors duration-150">
Read More
</button>
</div>
</div>{
"v": 1,
"tokens": {
"c": {"p": "#6366f1", "bg": "#1e293b", "fg": "#f8fafc", "mt": "#94a3b8", "pbg": "#6366f133", "ptx": "#818cf8"},
"sp": [0, 4, 8, 12, 16, 24, 32, 48],
"fs": [12, 14, 16, 18, 20, 24],
"r": [0, 4, 8, 12, 16],
"sh": {"lg": "0 10 15 -3 #0004", "xl": "0 20 25 -5 #0006"}
},
"tree": [
{
"id": "card", "maxw": 384, "r": 4, "ov": "hidden",
"sh": "lg", "bg": "bg", "tr": "all 300",
"c": [
{"tag": "img", "src": "/image.jpg", "alt": "AI-generated abstract visualization", "w": "100%", "h": 192, "fit": "cover"},
{
"id": "body", "p": 6,
"c": [
{"id": "badge", "txt": "Research", "bg": "pbg", "fc": "ptx", "fs": 0, "fw": 600, "px": 3, "py": 1, "r": 4, "d": "inline-block", "mb": 3},
{"tag": "h3", "txt": "The Human Overhead Thesis", "fs": 4, "fw": 700, "fc": "fg", "mb": 2},
{"tag": "p", "txt": "60–80% of AI-generated web code exists purely for human readability. We measured it.", "fc": "mt", "fs": 1, "lh": 1.6, "mb": 4},
{"id": "cta", "tag": "button", "txt": "Read More", "w": "100%", "bg": "p", "fc": "#fff", "fw": 500, "py": 3, "px": 4, "r": 3, "tr": "all 150", "cur": "pointer"}
]
}
]
}
],
"states": {
"card": {"hover": {"sh": "xl"}},
"cta": {"hover": {"bg": "#818cf8"}}
}
}The Human Overhead Thesis
60–80% of AI-generated web code exists purely for human readability. We measured it.
Identical. The authoring format changed. The output didn't.
The Numbers
Aggregate metrics across all seven test cases. Every measurement uses the same tokenizer (cl100k_base).
Tokens per Test Case
Cumulative Token Divergence
| Test Case | Traditional | Cost | CYPHIR | Cost | Reduction | A11y |
|---|---|---|---|---|---|---|
| Content Card | 187 | $0.0028 | 142 | $0.0021 | 24% | 100% |
| Responsive Nav | 312 | $0.0047 | 168 | $0.0025 | 46% | 100% |
| Data Table | 445 | $0.0067 | 198 | $0.0030 | 55% | 100% |
| Form + Validation | 523 | $0.0078 | 215 | $0.0032 | 59% | 100% |
| Dashboard Layout | 892 | $0.0134 | 324 | $0.0049 | 64% | 100% |
| Dark Mode Toggle | 634 | $0.0095 | 187 | $0.0028 | 70% | 100% |
| Animated Hero | 756 | $0.0113 | 267 | $0.0040 | 65% | 100% |
| Total | 3,749 | $0.0562 | 1,501 | $0.0225 |
Costs calculated at Claude Sonnet 4.6 output pricing ($15/M tokens)
Use CYPHIR
One compiler. Four ways in. Your tools, made better.
MCP Server
One command. Your AI starts speaking CYPHIR. Works with Claude Code, Cursor, VS Code, and any MCP-compatible client.
claude mcp add cyphir --transport http https://mcp.cyphir.ai
REST API
Embed CYPHIR directly into your platform's generation pipeline. Send IR, receive optimized web output.
POST api.cyphir.ai/compile
{ "ir": { "v": 1, "tree": [...] } }CLI
See what CYPHIR does to code you've already generated. Run it on any project and get a savings report.
npx cyphir analyze ./my-project
VS Code Extension
Find CYPHIR where you already work. Real-time compilation preview in your editor.
ext install cyphir.cyphir-vscode
Works with every LLM. Claude, GPT, Gemini, Llama — CYPHIR is model-agnostic.
The Architecture
CYPHIR removes the human-readable layer between AI intent and browser execution. Three stages. Zero runtime overhead.
AI Authors Intent
The AI generates a compact IR document — just structure, tokens, and relationships. No verbose syntax, no human-readable naming.
{ "id": "card", "maxw": 384,
"r": 4, "sh": "lg",
"c": [{ "tag": "h3",
"txt": "Title", "fw": 700 }] }Compiler Transforms
The CYPHIR compiler expands the IR into standards-compliant HTML, CSS, and JS — injecting accessibility, semantics, and best practices automatically.
IR → semantic HTML IR → optimized CSS IR → ARIA attributes IR → responsive rules IR → interaction handlers
Browser Renders
The browser receives standard web code — identical to hand-authored output. No runtime, no framework, no dependencies. Just HTML, CSS, and JS.
<div class="card" role="article" aria-label="Content card"> <h3>Title</h3> <style scoped>/* optimized */</style> </div>