--- title: I logged into 414 Chinese AI API resellers and priced every tier description: There is a large gray market in China reselling Claude and GPT API access for RMB, and no two of them price the same way. I logged into 414 of them, pulled all 3,159 pricing tiers, and normalized everything to one number — percent of official list price. Here is the method and what surprised me. published: 2026-08-09 updated: 2026-08-09 lang: en-US # 必须带地区：og:locale 要 language_TERRITORY 形式，光写 en 是无效值 alternate: pricing-audit keywords: - Chinese AI API resellers - API relay station - Claude API price China - 中转站 - LLM API pricing - gray market API --- # I logged into 414 Chinese AI API resellers and priced every tier I could not answer a question that should have been easy: if I put ¥100 into one of these resellers, what fraction of Anthropic's list price am I actually paying? Two months later I have an answer for 414 of them. This is how, and what I did not expect. --- ## Why this market exists If you are outside China, the obvious question is why anyone would buy API access from a third party instead of from Anthropic or OpenAI. The answer is usually not price. It is that the official door does not open. Four failure modes, roughly in order of how often I hear them: - **No card that works.** Chinese debit cards are not accepted; the UnionPay credit cards that are accepted often fail the issuer's own cross-border checks. Virtual card services get shut down in waves. - **No way to settle in USD.** A company can wire money abroad, but the paperwork for a $200/month API bill is not worth anyone's time. - **Reachability.** The endpoints are reachable from some networks and not others, and "sometimes" is not a basis for production. - **Invoices.** Chinese companies need a **fapiao** (official tax invoice) to expense anything. Anthropic does not issue one. So a market grew to sell what people cannot buy directly. The Chinese term is **中转站** — literally "transfer station", conventionally translated as **relay station**. There are hundreds of them. The one I have collected the most from has been running for over a year. ## What a relay station actually is Mechanically, it is unglamorous: a third-party endpoint that forwards your API calls upstream and bills you in RMB. ``` Official: your code ──► api.anthropic.com ──► model Relay: your code ──► relay /v1 endpoint ──► upstream ──► model ``` You change two environment variables — `ANTHROPIC_BASE_URL` and `ANTHROPIC_AUTH_TOKEN`, or the OpenAI equivalents — and nothing else. The wire protocol is unchanged. That compatibility is the entire technical basis of the industry. Where the upstream capacity comes from is generally not disclosed, and where it is disclosed it is not verifiable. What you see instead is the name the operator gave the tier: `aws-q`, `cc-sale`, "official relay", "reverse channel". Those are product names, not provenance. Most of these sites run the same handful of open-source billing gateways — `new-api`, its upstream `one-api`, `sub2api` — with a new logo and different numbers. Once you notice this, the vocabulary stops looking proprietary: "token", "group", "ratio" are that software's fixed terms, not any one operator's invention. ## Why you cannot compare them Every site sells **groups** — pricing tiers — each carrying a **multiplier** off official USD pricing. A group at `0.3x` bills you 30% of what Anthropic charges. Separately, each site sets its own **top-up rate**: how many RMB buy $1 of account credit. One site says ¥3.5 buys $1. Another says ¥5. Those are two independent discounts, and each site shows you only its own. There is no page anywhere, on any of these sites, that tells you how its price compares to the site next door. That is the actual gap — not that prices are hidden, but that they are stated in 414 mutually incomparable units. The fix is one line: ``` percent of official = group multiplier × top-up rate ÷ 6.8 ``` 6.8 is the real RMB/USD rate. Official list price is 100%. Every number on my site is that calculation, with denominators taken independently from the vendors' own pricing pages — never back-derived from a reseller's claims. ## The part that took the time **Anonymous listings are a shadow of the real ones.** This is the single thing that makes the work slow, and it is why nobody had done it. Fetching a site's public pricing endpoint gives you a plausible-looking list of tiers. Log in and you get a different list. On one site the anonymous view showed 22 groups and the authenticated view showed 24 — and one of the two missing groups was priced 40% below the cheapest one visible anonymously. I had already published a ranking based on the anonymous view. It was wrong by 67%. So the process for each site is: register, log in, read the authenticated group list, cross-check every tier against the site's own per-model pricing page, and record it. Four hundred and fourteen times. This is the step that cannot be batched, and it is why nobody had done it — not because it is hard, but because it has to be done one site at a time. The ones I could not finish fall into two groups, each logged with its reason. Some never show a price at all: human verification walls, invite codes, QQ-mail-only registration, phone number required. The others stall partway, usually because the site publishes a top-up bonus ("pay ¥100, get ¥120") but never publishes the number that converts it into USD credit. Two rules I learned by breaking them: - **The multiplier dropdown is not authoritative.** Many operators discount individual models and never update the group multiplier. When the per-model price and the group multiplier disagree, the per-model price is what you get charged. Two sites I compared came out nearly 2× apart depending on which I believed. - **A tier you cannot select does not count.** Some groups appear in a model's `enable_groups` list but are absent from a new account's group list — `Claude VIP`, `claude-max`, that kind of thing. Pricing those is inventing a discount that no new buyer can obtain. ## Three things I did not expect **1. Rankings flip between vendors.** I assumed a cheap site was a cheap site. It is not. One site in my current data sits 63rd of 351 for Claude and 341st of 389 for GPT. If you chose your provider by comparing Claude prices — which is what people do, because Claude Code is what they are running — you may be paying near the top of the market for GPT without any signal that this is happening. Inside any single site you cannot see this. It only appears across sites. **2. The same multiplier can mean two things 6.8× apart.** A group marked `0.8x` on a USD-priced model works out to 11.8% of official. The same `0.8x` on an RMB-priced model — Kimi, GLM, DeepSeek, which are natively priced in RMB — works out to 80%. The currency-conversion layer simply is not there for those models. Same string in the same dropdown, and if you apply one mental model to both you are wrong by a factor of 6.8. I got this wrong first and had to re-derive an entire vendor's table. **3. Tiers that cost more than official, and why I keep them.** Some tiers land above 100%. The obvious move is to drop them as noise. I do not, because "more than official" is only absurd if you can buy official. For someone with no card, no USD settlement, and an invoice requirement, a tier at 130% is not a joke — it is the only door. Dropping it would be me deciding on the reader's behalf whether it is worth it. I would rather print the number and let the gap speak. In a cross-site table those tiers sink to the bottom on their own, where anyone can see how much cheaper the alternatives are. ## What I deliberately do not measure Quality. Uptime. Whether the model you asked for is the model you got. Whether the operator will disappear with your balance next month. Not because those do not matter — they matter more than price — but because you cannot know them without buying access and running production traffic through it for months, on every site. That is a different project, with a different budget, and pretending to have done it would be worse than not doing it. So: **every number I publish is what the operator publishes.** It is a stated price, not a measured one. Where stated and actual prices diverge is its own subject, and I have written separately about the seven places that gap tends to open, with a check you can run yourself for each. ## Data The full dataset is open under CC-BY: [`/data.json`](/data.json) for structured data, [`/llms-full.txt`](/llms-full.txt) for all 3,159 tiers as a Markdown table, [`/llms.txt`](/llms.txt) for the summary. If you want to check my arithmetic, those are the inputs. **Disclosure:** a few of these operators pay a referral commission if you sign up through the site. Most do not. It has no effect on ordering — ordering is strictly ascending percent-of-official, and there is no commission field in the data model to sort by even if I wanted to. The full statement is on the [about page](/about). The site itself is in Chinese, because that is who needs it. So is [the Chinese version of this write-up](/guides/pricing-audit), which skips the background and goes straight to the method.