Blog / Developer tools

Paying for AI APIs in China: Invoices, Compliance and Gateways

2026-10-19cdnping editors9 min read

If you want to use an AI API without wrestling with a foreign-currency card, there are roughly three routes. You can pay in USD directly on the official platform with an overseas payment method. You can top up in RMB at a Chinese cloud provider or a domestic model platform. Or you can subscribe to an aggregation gateway that accepts RMB for AI API access. The three differ on invoices, compliance, data policy, and who you contact when something breaks. This post goes through each, and ends with a checklist to run before you pay.

#What are the three payment routes?

For the technical differences between these approaches, see Official API, cloud-hosted, or aggregation gateway: how to choose an access method. This post only covers how you pay, how you get an invoice, and how you stay within the rules.

#Official overseas API: the most direct, but payment and region are the hurdles

Using a vendor's own API means the models, billing, and terms of service all come from the primary source, and there is a single party responsible when something goes wrong. There are two hurdles:

1. Payment method. Overseas vendors usually price in USD, so you need a card that can pay in foreign currency.
2. Region restrictions. Each vendor lists supported countries and regions in its terms of service or documentation. Before you sign up, check the official docs to confirm your region is on the list. If it is not, do not take this route. Getting around region restrictions violates the terms of service, the account may be disabled, and the balance may not be recoverable.

If you are expensing this through a company, think one step further: ask finance up front whether a foreign-currency bill can be booked and whether they will accept it.

#Chinese cloud providers and domestic model platforms: easiest for RMB and invoices

Domestic model platforms and cloud providers settle in RMB, generally support enterprise verification and corporate payment, and issue invoices under Chinese rules (fapiao, the official tax invoice used for reimbursement and accounting in China). If your main work is Chinese long-text or batch processing, try your own tasks on a Chinese model first; if it is good enough, this route is the least hassle for finance and compliance.

Things to watch:

• Invoice type (regular or VAT special invoice), the invoicing entity, and timing all follow each platform's console documentation. Read them before you pay.
• API formats are not necessarily identical across platforms. Many offer an OpenAI-compatible interface, but parameter details can differ. Before switching platforms, run your own calls with a small balance.
• There is generally no cross-border network issue, but it is still worth measuring real latency.

#Aggregation gateways: one account for many models

An aggregation gateway puts models from several vendors behind one interface, usually compatible with the OpenAI or Anthropic API format, so a developer tool only needs a new endpoint URL and key. For individual developers, the draw is paying in RMB without an overseas card.

Services in this category vary a lot. At minimum, check four things: whether it is an officially authorized channel, how data is handled, whether billing is transparent, and whether uptime data is public. The next section covers these.

#What exactly should you check when choosing a gateway?

AGI Plan serves as the example for how to check item by item. First, the relationship: AGI Plan is a commercial partner of cdnping. So the list below only includes what AGI Plan states publicly on its own site, and every line can be verified by you.

The data policy is the item most often skipped. Before you hand company code to a third-party API, read the original text. For how to read it and which sections matter, see Can your code and conversations be seen when you use a third-party AI API?. For a fuller set of relay and gateway evaluation criteria, see How to choose an AI API relay: 8 checks on uptime, billing, and data policy.

#How do you convert subscription quota into usage?

AGI Plan's pricing is on its site: Lite ¥49/month, Plus ¥119/month, Pro ¥649/month, Max ¥1,368/month. The four tiers differ only in quota, and annual billing is 20% off. The unit of measure is the CU: 1 CU equals 1,000 output tokens of the baseline model. Pay-as-you-go is also available at ¥1 = 500 CU, with a ¥20 minimum top-up.

By that definition you can do the math yourself: a ¥20 top-up gives 10,000 CU, equivalent to 10 million output tokens of the baseline model. Two caveats:

• The conversion applies to output tokens of the "baseline model." Other models burn quota at different rates, so how much you actually get depends on which capability group you call.
• Input tokens and output tokens are counted separately. If the month-end bill does not match your local logs, check the counting method first and the quantity second. See AI API bill doesn't match: token counting methods and how to reconcile.

A sensible approach: start with the smallest pay-as-you-go top-up, run your real workload for a week, note the daily CU consumption, then decide which subscription tier to buy.

#How do you verify uptime yourself?

Whichever provider you pick, any uptime figure it gives should trace back to raw data. AGI Plan publishes hourly, per-capability-group success and failure counts, with no login required:

# Returns CSV by default
curl -s "https://agiplan.dev/api/status/requests?month=2026-10"

# Add format=json for JSON, handy for summarizing with jq
curl -s "https://agiplan.dev/api/status/requests?month=2026-10&format=json" | head -c 1000

The definition is: uptime = successes / (successes + failures caused by the platform itself), and only 5xx counts as a platform failure. In other words, your own 4xx errors (wrong key, bad parameters, quota exhausted) are not counted.

The table below is data cdnping pulled from this public endpoint on October 17, 2026. The window is the 30 days up to that day, using the same definition.

The fast group is clearly lower, and it is listed here as-is. Capability groups differ a lot, and looking only at the total hides problems. If most of your load lands on one group, focus on that group and log your own failure rate on the client. cdnping's AGI Plan status page mirrors this data. To measure other services with the same definition, see Measure AI API uptime yourself: metrics, scripts, and definitions.

#What should you ask about invoices and reimbursement up front?

Invoices are where team purchases most often get stuck. Before paying, send the questions below to the provider and get written answers first:

1. What type of invoice can they issue, regular or VAT special?
2. Which company is the invoicing entity, and does it match the payee?
3. Can both subscriptions and pay-as-you-go top-ups be invoiced, and is the invoice amount based on the top-up or on consumption?
4. Can annual billing be invoiced in one go?

Official overseas APIs generally issue foreign bills, and whether they can be booked is a question for your own company's finance team. Invoicing rules differ across domestic platforms and gateways; go by the written statement on their site or from support, and do not assume.

#What are the compliance baselines?

This part is not complicated, but every item has to hold:

• Terms of service. Whichever route you take, follow the terms of the model vendor and of any intermediary platform. In regions the vendor does not support, do not use the service by faking location information, bulk registration, rotating multiple accounts, and the like. Those practices violate the terms and are out of scope for this post.
• Filing when you offer a service to others. Calling a model from your own code while you write software is one thing. Packaging model capability into a product and offering generative AI services to the Chinese public is another. The latter falls under filing or registration requirements in rules such as the Interim Measures for the Management of Generative AI Services. Consult legal counsel or the local regulator before launch.
• Cross-border data and sensitive data. Before sending company code or customer data to an overseas endpoint, check your internal data compliance requirements. Anonymize whatever you can.
• Website acceleration has its own requirements. If you want to accelerate an AI app for users in mainland China from inside the mainland, you need an ICP filing (the license required to operate a website in China) and must go through a licensed provider.

#Pre-payment checklist

Print this table or copy it into your notes and tick items off one by one:

A note on item 7: cdnping's /status probes from Cloudflare's edge. The October 17, 2026 snapshot came from the Denver (DEN) node, with 242 probes in total. It reflects the time from an overseas data center to each API entry point, not your actual experience inside China. For latency on a Chinese network, use the home page speed test from your own machine.

If item 8 gets stuck, see Claude Code custom API, and endpoint configuration for Codex and OpenCode for each tool.

#How should you choose?

A rough split by use case:

• Company purchasing, special VAT invoices needed, mainly Chinese-language business: start with Chinese cloud providers and domestic model platforms.
• Individual developers writing code day to day, wanting to switch between several models in one tool, with no overseas card: consider a gateway that bills in RMB, but run through the checklist above, especially "officially authorized or not" and the data policy.
• A firm requirement on model provenance, your region is within the official supported range, and you can handle foreign-currency bills: use the official API directly.

You can also mix them, for example production traffic through the official API or a cloud provider, and development and experiments through a gateway. Whatever the combination, trial it with a small amount for a week, check uptime and billing against your own logs, and only then commit to long-term payment.