// the find
dimthink/PriceAI
AI 订阅卡网渠道比价工具：聚合100+卡网渠道包含 ChatGPT、Claude、Gemini、Grok 等多渠道报价，展示有货最低价、库存状态和原站购买链接。
PriceAI is a price aggregator for AI subscription channels in the Chinese market — it tracks 200+ third-party card shops selling ChatGPT Plus, Claude Pro, Gemini, etc. at below-official prices, alongside official regional pricing and API relay station comparisons. If you're a Chinese developer trying to figure out whether to pay ¥200 on a card shop or just use an official API transit, this is the tool for that decision.
The four-module split (official prices, card-shop subscriptions, official APIs, transit APIs) is a real design choice — each category has different risk profiles and verification requirements, and keeping them separate prevents users from accidentally treating a gray-market card shop the same as an official API. The GitHub Actions pipeline has distinct workflows per data source (dujiao, ldxp, official prices, API transit probes) rather than one monolithic scraper, so a broken collector degrades gracefully instead of taking everything down. The standardized product/model/station abstraction is doing the hard work: normalizing '成品号 ChatGPT Plus' and 'CDK Plus 代充' into the same comparison row requires real domain modeling. The license strategy — source-visible but no commercial deployment — is explicitly protecting their moat without pretending to be AGPL.
The transit API (中转 API) section is tracking resellers who may or may not be reselling stolen API keys; the codebase has no mechanism to flag this, and publishing uptime stats implicitly legitimizes services with opaque upstream provenance. Stale data is the core reliability problem here — the collection scripts run on timers and if a scraper breaks, there's no visible staleness indicator at the UI layer, so a user could be looking at week-old price data with no warning. The Supabase dependency is a single point of failure for both the write path (price ingestion) and the read path (public site), with no caching layer between them on the application side — a Cloudflare-deployed Next.js app hitting Supabase on every request at scale will hit connection limits before it hits traffic limits. There are no tests for the collector scripts, which are the most business-critical code in the repo.