relayapi

This is a review and recommendation document for AI API relay stations aimed at Chinese users. Based on personal experience, the author categorizes relay stations encountered into three groups: "recommended," "neutral," and "not recommended," and outlines several dimensions to consider when choosing a relay station. Regarding selection criteria, the author emphasizes: stability first, preferring stations with good reputations, long operating histories, and active user communities; second is response speed; third is model coverage, ideally offering one-stop access to mainstream models and quickly following up on official new models; fourth is billing issues, including token counting methods, rates, and bill clarity, while warning that suspiciously cheap models may be low-cost domestic models masquerading as expensive foreign ones; fifth is the risk of operators disappearing, prioritizing company-operated stations and monitoring customer service responsiveness and rate fluctuation notices; sixth is hidden costs, such as cache pricing and cache hit rates, with normal cache pricing around 10%, while some stations charge 15% or even 30%. The author repeatedly advises against large recharges, recommending topping up only as much as needed. The "recommended" section lists multiple stations, including CUN.ai, Modelflare, hao.ai, SudoCode, EiRouter, 9527code, 灵算, ToTokens, ToolCode, 云渡, CheapAI, znbcode, DuiAPI, 智惠API, Portdan AI, XycAi (星道智能), 我的贾维斯, YKH.AI, ModCon, and others. Each entry roughly describes launch time, supported models (such as GPT, Claude, Gemini, Grok, DeepSeek, Qwen, GLM, etc.), price examples (expressed in RMB per million tokens for input/output), payment and invoicing capabilities, refund policies, group types (such as special-offer groups, stable groups, Pro groups), and promotions like free credits for new users. The "neutral" section includes Chintao AI, RightCode, aigocode.com, OpenRouter.ai, 哈基米, Ekan8, and others. The author records their interface quality, pricing, and stability changes, noting issues such as diluted models, higher-than-expected token consumption, or interface fluctuations. The "not recommended" section includes relay stations on Xianyu (second-hand marketplace), AI派, 大肘子, and others, with main problems centered on model dilution, subpar speed, excessive cache pricing, opaque billing, slow interfaces, or the risk of operators disappearing. The document also mentions the author's maintained 禾维 AI site, which offers relay station rankings, port detection, and real-time price updates, and welcomes relay station operators to submit review information. Overall, the content serves as experiential evaluation and purchasing reference, with prices and promotional information changing rapidly over time.