Pay for Claude, GPT and Gemini APIs With Alipay or WeChat Pay (CNY): Top-Up vs API Billing, and How to Check Each
Paying for an AI API with Alipay or WeChat Pay is two separate steps: funding a balance, then spending it per token. How each step works, what can go wrong in each, and a quick self-check to tell which one failed.
1. Two steps, two different failure modes
When diagnosing a prepaid AI API, separate funding from spending. Funding is a payment: money moves from your Alipay, WeChat Pay or card into a prepaid balance. Spending is metering: each request reports how many input and output tokens it used, the price per million tokens is applied, and the result is deducted from that balance. Anthropic describes its own model in exactly these terms — Claude API usage is billed through prepaid usage credits, bought before you use the API and applied to usage at current pricing. Keeping the two steps apart in your head saves hours. A payment that did not arrive is a question for the payment record, not for your code. An API call that fails while the balance is healthy is a question for your key, account status, model name or request, not for Alipay. The table above maps the common symptoms to the step they belong to; the sections below go through each one and give you a check you can run yourself before contacting anyone.
2. Where official consoles get stuck: the funding step
The official consoles meter usage well; the hard part for many readers is the checkout. Anthropic's help center sends you to the Console's Billing page to buy credits and to an update-payment-method form to change your card; it notes that credits expire one year after purchase, that the expiry cannot be extended, and that all credit purchases are non-refundable. It also offers auto-reload, which buys more credits when your balance drops below a threshold you set. OpenAI and Google run their own billing pages; check each one's billing settings and help center for the payment methods it accepts today rather than relying on an old forum post. If your locally issued card is accepted, the official route is the most direct one: no intermediary; check its current model access and settings before choosing. If the checkout keeps declining your card, the reason is usually on the payment side — how the issuer and the merchant's processor score that transaction — not on the API side, and retrying the same card rarely changes the outcome. See /hub/why-china-card-declined for how those declines work.
3. How a CNY top-up works on a gateway
A gateway that takes local payment methods replaces only the funding step. On cocodot the flow is: register, open the wallet, choose a top-up method — Alipay / WeChat Pay (RMB), USDT / USDC — and pay. The wallet is denominated in US dollars; a fiat top-up is converted and credited at the exchange rate locked in when you place the order, so what you see on the confirmation is what lands in the wallet. The current minimum amount and any method-specific notes are shown on the top-up page itself; check there rather than in articles. After that, the spending step works like any OpenAI-compatible API: create a key in the console, point your client at https://cocodot.co/api/ai/v1, and choose a model. The same wallet balance pays for Claude, GPT and Gemini models, plus several Chinese model families, so you do not keep one prepaid balance per vendor. Listed Claude / GPT / Gemini rates are at or below the official list price; check the pricing page for each model's current rate.
No overseas card needed: cocodot's OpenAI-compatible API — switch your base_url to call the models on sale.
Top up via Alipay / WeChat Pay (RMB), USDT / USDC · billed per token · switch your base_url to connect
4. Self-check for the funding step
If you paid and the balance did not change, work through this before anything else. (1) Look at the top-up record in your wallet, not only at the Alipay or WeChat Pay receipt. The receipt proves money left your account; the wallet record shows whether the order was matched and credited. (2) Do not pay again to "push" the first payment through. A second payment is a second order; it does not speed up the first and leaves you two orders to reconcile. (3) Keep the order details — time, amount and the payment app's transaction number — so support can match it in one message. (4) Compare currencies correctly. You paid in RMB and the wallet shows USD at the rate locked for that order, so the two numbers will not look alike; check the conversion on the order, not against today's rate on a search engine. (5) Remember the wallet and a virtual card are separate balances. Money in the wallet does not appear on a card until you load it, and loading a card from the wallet costs 3%. For API calls you never need a card: the wallet pays for them directly.
5. Self-check for the spending step
If the wallet shows money but calls fail, the problem is in the request path, and the status code points to the cause. 403 on every call: check whether the account has finished the identity check (real name plus an email code, about a minute, free); the API stays closed until it is done, whatever the balance. 401: the key is wrong, truncated, revoked or sent in the wrong header; send it as `Authorization: Bearer <key>` and check the configured key locally without printing or logging any part of it. Model not found: compare your model string with the public list at https://cocodot.co/api/ai/v1/models (no key needed) and copy the name exactly. 402 asking for a top-up: the account has not made its first top-up yet; the trial credit is usable after it. A flagship refuses while a small model works: the wallet is at zero and only trial credit is left, which covers selected budget models only; Claude, GPT and Gemini flagships need a wallet balance above zero. Calls work but the answer is cut off: that is a `max_tokens` or context question, not a payment one; see /hub/context-length-exceeded-fix. Key errors in more depth: /hub/api-401-invalid-api-key-fix.
6. Reading the bill: where the balance actually goes
Once calls work, the next surprise is usually cost. Price a request rather than a month: input tokens times the input rate plus output tokens times the output rate, each per million. Three things inflate the bill without changing what you see on screen. Conversation history: most chat clients resend the whole conversation on every turn, so turn twenty pays for turns one to nineteen again as input. Reasoning tokens: check the model's current official billing documentation for how reasoning tokens are metered; visible answer length alone may not describe the usage shown in `usage`. Retries: a retry loop that fires on every error multiplies spend when the cause is something a retry cannot fix. A quick audit: pick ten real requests, log `usage` for each, and add them up by hand. If the total matches the console, your metering is right and the cost is the workload itself; if it does not, look for hidden retries or duplicated history. Prompt caching changes the input side: cache hits at half the input price; Claude cache writes at 1.25x the input price (2x for the portion written to the 1-hour cache), the same structure as the official pricing.
7. Choosing between the official console and a CNY gateway
Use the official console when its checkout accepts your card and you want the direct relationship: nobody between you and the model. Check the official console for the models, settings and account controls available to your account today. Use a gateway when the funding step is the blocker, or when you want one balance across several vendors. Be honest about the trade: your counterparty becomes the gateway operator, and you carry the risk of a prepaid balance with them, so read their terms on balances and account suspension before you top up a large amount. Check the model too. Do not infer model identity from a display name alone; an open-source probe (probe.cocodot.co) sends a fixed prompt set to any OpenAI-compatible endpoint and scores whether the answers fit the claimed model, and you can run it against any provider, including the one that publishes it. For a broader buyer's checklist see /hub/how-to-choose-ai-api-relay.
8. A first top-up that tells you everything
Treat the first payment as a test, not a commitment. Top up the smallest amount the top-up page allows, finish the identity check, create one key, and run three calls: a small model, the flagship you actually plan to use, and the same flagship with a long prompt that resembles real work. Check three things after each call: the response, the `usage` numbers, and the matching deduction in the console. If all three line up, you have confirmed both steps — funding and spending — with very little at stake, and you can raise the amount knowing what a typical request costs you. If something does not line up, you now know which step to look at, and you have a small, specific case to send to support instead of a vague "it doesn't work". Developers moving an existing project over should also check SDK compatibility: most OpenAI-SDK code needs only a new base_url and key, and Claude-native code can use the Anthropic-compatible endpoint described in /hub/anthropic-api.
Funding vs spending: symptom → which step → what to check
| What you see | Which step | What to check first |
|---|---|---|
| Paid in Alipay, wallet balance unchanged | Funding | The top-up record in your wallet — do not pay a second time |
| Balance shows, but every call returns 403 | Spending — account not cleared for API yet | Identity check (real name + email code) completed? |
| 401 or "invalid API key" | Spending — credentials | Key copied whole, sent as Bearer, right base_url |
| Trial credit works on a small model, a flagship call is refused | Spending — balance type | Wallet balance above zero; leftover trial credit alone covers budget models only |
| Model not found | Spending — model name | Name or code copied from the public models list |
| Bill higher than expected | Spending — usage | Token counts per request, retries, long histories |