Every row below is a real page read through the Lyrenth index, with the token economics reported by the API itself: what the raw HTML would cost a model to ingest, what the clean AIDocument costs instead, and the difference. Measured 2026-07-05. Original data, reproducible with one API call.
| Page | Raw HTML tokens | AIDocument tokens | Smaller | Saved | First read | Repeat read (median) |
|---|---|---|---|---|---|---|
| Stripe API reference Payments | 307,902 | 2,000 | 154x | 99.4% | 959 ms | 4 ms |
| Vercel Functions docs Serverless | 237,390 | 2,149 | 110x | 99.1% | 171 ms | 2 ms |
| Cloudflare Workers docs Serverless | 94,963 | 1,377 | 69x | 98.5% | 128 ms | 3 ms |
| GitHub REST API quickstart Developer platform | 88,614 | 4,876 | 18x | 94.5% | 572 ms | 3 ms |
| Kubernetes Pods concepts Infrastructure | 131,977 | 7,475 | 18x | 94.3% | 375 ms | 4 ms |
| React useState reference Frontend | 111,042 | 8,156 | 14x | 92.7% | 213 ms | 8 ms |
| MDN: Overview of HTTP Web fundamentals | 56,692 | 5,085 | 11x | 91.0% | 677 ms | 5 ms |
| Python asyncio docs Languages | 6,175 | 1,251 | 4.9x | 79.7% | 196 ms | 3 ms |
| Node.js fs module Languages | 259,563 | 102,372 | 2.5x | 60.6% | 446 ms | 10 ms |
| PostgreSQL SELECT reference Databases | 29,509 | 17,500 | 1.7x | 40.7% | 439 ms | 4 ms |
Cost basis: $3.00 per 1M input tokens (mid-tier frontier model class), as reported in the API's economics block.
The last two columns are the same ten pages, measured 2026-08-06: what each cost to read before it was indexed, and after. Every repeat read came back in 10 ms or less, whatever the page weighs.
A scraper fetches the page every time you ask for it, so it pays the open web's latency on every call. An index pays it once. The numbers below are percentiles over real reads through the production API, measured 2026-08-06.
| Read path | Reads | Median | 95th pct |
|---|---|---|---|
Served from the index The page is already an AIDocument. No origin contact at all. | 714 | 12 ms | 126 ms |
First read, server-rendered page Not indexed yet: fetched live, parsed, and stored on the way out. | 1,405 | 386 ms | 1.96 s |
First read, JavaScript-only page The page ships an empty shell, so a real browser has to run its JavaScript before there is anything to read. | 211 | 2.62 s | 6.27 s |
The bottom row is the one worth reading twice. A JavaScript-only page is the slowest thing on the web to read, and we take that hit once, on our side. Every read after it is served from the index at a median of 32 ms. That is the difference between fetching a page and having indexed it.
Percentiles over every successful read in the production request log, server-side handler time. Your own round trip adds the network between you and us, so measure it from where your agent runs.