Sinan Compute · 司南·算力

Effective-price comparison and reachability measurement for model API relay sites. 中转站实付比价与可达性测量 —— 每个数字都能点开抓取快照。

🌐 https://compute.sinanlab.com · 🇬🇧 https://compute.sinanlab.com/en/ · 📡 RSS · 🧾 Methodology · 📦 Raw data

📈 Token Price Index / Token 价格指数	market median effective price ÷ official price, three tiers, daily · page · JSON
🏆 Sinan Rankings / 司南榜	12 measured boards, weekly, permanent links · page
📰 Monthly report / 月报	market size, structure, prices, reachability, probes, multimodal, compute · page · Releases
🧪 sinan-probe CLI	test a relay with your own key, stdlib only · cli/
🤖 GitHub Action · price watch	copy one workflow, get an Issue when your models' relay-market price moves · templates/github-action
📚 Open data & API docs	every file, field, update cadence and the stability promise · page
🛠 Corrections log	user reports, verification results, credited · page
✅ Verified by Sinan / 经司南核验	free, criteria public, auto-granted and auto-revoked · apply
📣 Press & researchers / 媒体与研究者	citable data and formats · page

Data files in site/reports/, site/rank/, site/weekly/ are committed nightly by the pipeline.

---

What it does / 这是什么

Hundreds of "relay sites" (中转站) resell LLM APIs at a discount. Their panels show a nominal USD price, but what you actually pay depends on a top-up rate, currency and multiplier that differ per site. Sinan Compute turns all of that into one number:

effective $/M tokens = panel nominal price × top-up rate (¥ per $1 nominal) ÷ USD/CNY
ratio                = effective ÷ lowest public reference (official page or OpenRouter)

Every ratio is then placed in an arithmetic band (below cost floor / below bulk discount / within range / near public price / above). We state the band; we do not speculate on why.

Relay sites indexed	900+ (confirmed by panel fingerprint; 300+ with public pricing)
Models tracked	40 text models + 27 image/video families
Quotes	14,000+ effective prices, refreshed nightly
Reachability	hourly probe, 24h uptime + p50 latency per site
Consistency probe	12 fixed prompts per model, token-count fingerprint compared across sites
Evidence	every number links to a sha256-hashed fetch snapshot

What it is not / 我们不做什么

• No recommendations, no rankings by "quality", no "scam" labels. Ratios are arithmetic; wording is gated by a banned-word linter before every deploy.
• No money from any measured vendor: zero affiliate links, zero ads, zero sponsors (why trust us).
• Not a verdict on authenticity. The consistency probe reports counts consistent / counts differ / too few peers, nothing more.

Use the data / 用数据

Endpoint	Content
/data_v2.json	model ledger, 900+ sites, snapshot index, FX rate
/media.json	image / video quotes vs official prices
/history/<model>.json	price history per model
/feed.xml	daily price changes + new sites
/citation-context.json · .md	compact, timestamped context for AI assistants and researchers
/llms.txt	how to read and cite this site
/badge/<domain>.svg	embeddable measured-facts badge for site owners

License for data: CC BY 4.0 — cite compute.sinanlab.com and the snapshot date.

How it works / 怎么运转

CT-log tail + directory harvest ──▶ panel fingerprint (/api/status, /v1/models)
        │                                        │
        ▼                                        ▼
  relay_candidate (SQLite)  ──▶ pricing adapters (one-api / new-api …) ──▶ offer_norm
        │                                        │
  hourly heartbeat (uptime, TTFB)          official reference pullers (OpenAI, Anthropic, Google, DeepSeek,
        │                                  Volcengine, Aliyun Bailian, Vidu, Kling … + OpenRouter)
        ▼                                        ▼
  export_data.py ──▶ data_v2.json ──▶ build_v5.py (static HTML, zh + /en/) ──▶ Cloudflare Pages
                                   └▶ i18n_apply.py · seo_assets.py · citation_context.py · notify_email.py

Everything runs once a night (03:00 Beijing). No framework: Python 3.9 + httpx, one static-site generator, Cloudflare Pages Functions for login / watch / mail.

Run it yourself / 本地运行

pip install httpx pyyaml pillow markdown
python3 site/export_data.py && python3 site/export_media.py && python3 site/build_v5.py
python3 -m http.server -d site/dist 8792

Data lives in data/compass.sqlite (not in the repo) and is produced by the collection engine (private). The generators above run against the exported data_v2.json / media.json, which are also published at the URLs listed under Use the data.

Corrections / 纠错

Found a wrong number? Open an issue with the page URL and a source link, or email [email protected]. Corrections are logged publicly; original records are kept.

Layout / 目录

Path	What
site/build_v5.py	static site generator (zh) · site/i18n_apply.py + site/i18n_dict.py → /en/
site/export_data.py	SQLite → data_v2.json (bands, clusters, probes)
collection engine	pricing adapters, discovery, probes and the nightly runner live in a private repo; this public repo holds the presentation layer, the user system, the methodology and the data snapshots
functions/	Cloudflare Pages Functions: GitHub login, watches, email subscriptions, alerts
docs/METHOD.md	the full methodology, versioned

---

Part of Sinan Lab / 司南实验室 — neutral measurement for AI infrastructure. Sister project: Sinan Robo, an auditable index of open embodied (VLA) models.