How Companies in China Buy Claude and GPT API for Dev Teams

Companies in mainland China that want Claude or GPT for their developers usually pick one of five routes: personal subscriptions that the company reimburses, a direct contract with an overseas cloud or the model vendor's own API, Chinese models on a domestic cloud platform, a relay station, or an API gateway. Start with what finance needs. If the books require Chinese VAT invoices (fapiao) and payment by corporate bank transfer, use Chinese models on a domestic platform, or a hybrid of those models plus a gateway. If the company has an entity outside the mainland (Hong Kong included), or finance accepts foreign invoices or receipts, and the team needs Claude, GPT and Gemini with a spend cap per developer, use an API gateway. Personal subscriptions only work for a handful of people. Below, the five routes are compared on payment, proof for finance, stability and traceability, and per-developer spend control.

Facts as of 2026-10-23

• Neither mainland China nor Hong Kong is on the supported-regions list for OpenAI's or Anthropic's own API (both public pages, checked 2026-10-23).
• In this post, the "domestic cloud platform" route means Chinese models on the model platforms of Chinese cloud vendors, not Claude or GPT; corporate bank transfer and fapiao follow each platform's own terms.
• Router One payment: Top up with a card or Alipay through one hosted checkout, or with USDT/USDC on six chains (Tron, BSC, Ethereum, Polygon, Base, Arbitrum). No US credit card required. Every hosted-checkout payment is invoiced automatically by email once it succeeds; these are not Chinese VAT invoices (fapiao), and stablecoin top-ups are not invoiced. Router One does not offer corporate bank transfer, purchase orders or payment terms.
• Router One contracting party: this site does not list a contracting company today; for an annual contract, email support@router.one with the subject "Annual contract".
• Router One keys: A key's spend cap (maxSpend) limits what that key can charge to your wallet: once its cumulative wallet charges reach the cap, its wallet-billed requests return HTTP 402. Requests covered by your plan's quota draw from the account's shared plan quota and do not count toward any key's cap.

An API gateway such as Router One fits if your company has an entity outside the mainland (Hong Kong included), or finance accepts foreign invoices and receipts; your developers need several model families — Claude, GPT, Gemini — including coding tools such as Claude Code and Codex; and you want a spend cap per developer and a monthly breakdown.

Look elsewhere if finance accepts only Chinese VAT invoices, or you must pay by corporate bank transfer, purchase order or on payment terms; you must sign a contract before a trial; you need SSO, member accounts and roles; or Chinese models already cover your work.

Paying as an individual developer? See How to Pay for LLM APIs Without a US Credit Card. This post is about a company buying for a whole dev team.

Ask three questions first: who pays, what finance needs, who signs off

The same three questions keep coming up on V2EX, the Chinese developer forum (view counts as of 2026-10-23):

• Who pays. In "请教一下各位，你们平时用 AI 是公司报销还是自费？" ("Does your company reimburse your AI use, or do you pay yourself?", t/1230423, July 2026, about 4,000 views), more than half of the 34 replies say the company pays in full or in part.
• Whether reimbursed personal accounts last. In "国内公司如何采购 Claude Code？" ("How does a company in China buy Claude Code?", t/1200488, March 2026, about 3,900 views), the poster says the company plans to buy around 50 accounts; today staff use their own accounts and get reimbursed, and the accounts "get banned easily and are unstable" (很容易封号，不稳定).
• What finance and the boss accept. In "codex 怎么官方购买按量付费版-企业付费" ("How to buy Codex pay-as-you-go officially, paid by the company", t/1210858, May 2026, about 3,600 views), the poster says the company requires buying from the official site, but he has no overseas card and no phone number; when someone suggested a relay station, he replied that relays won't do and "the boss won't agree" (老板不同意). In "关于现阶段采购 AI 的一些问题咨询" ("Questions about buying AI right now", t/1236950, August 2026, about 2,300 views), the team has about 100 developers, runs mostly on relay stations today (主力中转) and spends "no more than 10W" (100,000) a month; asked why not stay on relays, the poster points to the company's information-security work and writes: "never trust any relay station" (永远不要相信任何一个中转站).

For your own company, answer these three first:

1. Who pays: the company centrally, or staff who pay and claim it back?
2. What proof finance needs: Chinese VAT invoices and corporate bank transfer, or will foreign invoices and receipts do?
3. Who signs off, and who can veto: the boss, finance, security — and what each will not accept.

With those answers, the table below rules out most routes for you.

The five routes compared (as of 2026-10-23)

What to watch on each route

1. Personal subscriptions + reimbursement

The quickest option: staff subscribe to a personal Claude or ChatGPT plan and the company reimburses the receipts. The problem is the word "personal". The accounts are in employees' names, and signing up and renewing need an overseas payment method; mainland China is not on either vendor's supported list, and the t/1200488 poster says these accounts "get banned easily". When someone leaves, the account and its conversation history usually leave with them. Usage is limited by the personal plan, so going over means upgrading or waiting for the allowance to come back. Fine for a few people; not a way to equip a whole dev team.

2. Overseas cloud or the vendor's own API, direct

The route with the most complete paperwork: contract directly with the model vendor or an overseas cloud, get the vendor's bills and invoices, and use the vendor's own usage console. The hurdle is opening the account: the vendors' own APIs serve supported regions only (see the facts box above), and overseas clouds review access to their model services under their own regions and terms, usually requiring a company entity and payment method in a supported region. t/1210858 is the typical case: the company insists on the official site, and the team has no overseas card or phone number. If your overseas entity passes onboarding, this is the least hassle; if not, read on. For region errors such as unsupported_country_region_territory, the causes and fixes are on the region-block guide.

3. Chinese models on a domestic cloud platform

The route finance finds easiest: the model platforms of Chinese cloud vendors usually take RMB by corporate bank transfer and issue Chinese VAT invoices; check each platform's own terms. The cost is the models: this route means Chinese models, not Claude or GPT themselves. Chinese models are good enough for many tasks — see how Qwen and Doubao compare with Claude and GPT — but if your developers use Claude and GPT inside Claude Code and Codex, this route cannot provide that.

4. Relay stations

Usually the lowest price and the fastest start; the problems sit where you cannot see them. Is the model answering you the one you paid for? If the site shuts down one day, is your balance still there? When something breaks, is there a per-request record to reconcile against? To check model authenticity and spot downgrades yourself, see the LLM API authenticity guide; for the questions to put to any provider before production traffic, see LLM gateway vs API relay. Relays suit personal trials and demos, not company workloads or finance reconciliation.

5. API gateways

A gateway puts many models behind one OpenAI-compatible endpoint: the company tops up one prepaid wallet, each call is charged at posted rates, and each request sent to a model leaves a log row. For a company purchase, three things matter: a payment method that works from China, one key per developer with its own cap, and a monthly per-developer spend breakdown for finance. Router One's row uses only facts this site publishes:

• Payment: Top up with a card or Alipay through one hosted checkout, or with USDT/USDC on six chains (Tron, BSC, Ethereum, Polygon, Base, Arbitrum). No US credit card required.
• Invoices: every hosted-checkout payment is invoiced automatically by email once it succeeds, with nothing to request; these are not Chinese VAT invoices (fapiao); stablecoin top-ups are not invoiced.
• Corporate transfer: Router One does not offer corporate bank transfer, purchase orders or payment terms.
• Contracting party: this site does not list a contracting company today; for an annual contract, email support@router.one with the subject "Annual contract".
• Per-developer spend control: an account can hold many keys; give each developer one, each with its own spend cap. A key's spend cap (maxSpend) limits what that key can charge to your wallet: once its cumulative wallet charges reach the cap, its wallet-billed requests return HTTP 402. Requests covered by your plan's quota draw from the account's shared plan quota and do not count toward any key's cap.
• Reconciliation: Dashboard → Usage ranks your keys by spend, tokens and requests for the selected period; Dashboard → Logs filters by key and exports the filtered rows to CSV (request ID, model, key, tokens, cost, total time, status), which works as the per-developer spend record.
• Stability evidence: requests are served by the model you name, and a route that fails with 429, 5xx or a timeout is retried on another route for the same model; the model pages publish each model's 7-day and 30-day availability, and the status page lists active incidents; a written SLA with credits is available only under an annual enterprise agreement.
• Prices and refunds: billing follows the posted rates in the model catalog, and material price increases are announced at least 7 days ahead (see pricing methodology); purchased credits are non-refundable once added to your balance, except for billing errors on our side or where required by law (see the refund policy).
• Terms: calls through Router One are governed by Router One's terms of service. Keys may be issued to your own employees and contractors working on your behalf; keys stay under your control, are never published and are not handed to third parties for their own use; the account holder is responsible for all usage on the account.
• Data: for how request data is handled, see data retention.

For the full written answers on invoices, payment, contracts, limits and refunds, see Router One for companies — it is written to be forwarded to finance and legal.

The hybrid: Chinese models on a domestic platform, Claude and GPT through a gateway

Many teams end up using both:

• Usage that must be booked against Chinese VAT invoices — an internal knowledge base, customer service, bulk text processing — runs on Chinese models on a domestic cloud platform, paid by corporate transfer.
• Developers' coding tools, and any task that needs Claude, GPT or Gemini, go through the gateway: one key per developer, a cap per developer, a monthly CSV export.
• Which data may go to models outside China is your company's own data-security call.
• Keep two ledgers: the domestic platform's bills and invoices on one side; on the gateway side, a monthly Logs CSV summed by key, one key per person.

Rolling it out to a team of about ten

The short version is below; the step-by-step for Claude Code and Codex is in rolling out Claude Code and Codex to a dev team.

1. Register one account with a company email and give one person the job of topping up; at checkout, enter the email address that should receive the invoices.
2. Create one key per developer, named after the person, with a spend cap; give contractors' keys an expiry date.
3. For setup, see Claude Code in China and Codex CLI in China.
4. Each week, check the by-key ranking in Dashboard → Usage; at month end, export a CSV per key from Dashboard → Logs for finance. For per-key bookkeeping, see LLM Cost Attribution: One Key per App, Agent, Environment and LLM cost tracking.
5. When someone leaves, delete their key the same day. Expired keys that are not deleted still count toward the account's active-key limit, so do not wait for a key to expire.
6. If the account has a plan: plan quota is shared by the whole account and does not count toward any key's cap, so one developer's agent can use up the plan quota for everyone; to stop one person completely, delete their key.

FAQ

What do we do with a key when an employee leaves? Delete it in Dashboard → API Keys. Before deleting, you can export that key's rows from Dashboard → Logs to CSV for your records. Expired keys that are not deleted still count toward the account's active-key limit, so delete rather than waiting for expiry.

How much does each developer spend per month? It varies widely. In the V2EX thread "大家一年的 token 费用大概预估在多少钱？" ("Roughly how much do you spend on tokens a year?", t/1238355, August 2026, about 8,200 views and 85 replies), most personal figures fall between about 100 and 1,400 yuan a month, with one reply at more than 80,000 yuan a year. Companies elsewhere find it just as hard to forecast: according to The Information and Bloomberg in 2026, Uber used up its AI budget for all of 2026 in about four months and then capped spend at $1,500 per person, per tool, per month. Start each key on a conservative cap, watch Dashboard → Usage for a week or two, then adjust.

What is wrong with personal subscriptions that the company reimburses? The accounts sit in employees' names, signing up and renewing need an overseas payment method, and mainland China is not on the vendors' supported lists; on V2EX, companies report these accounts getting banned (t/1200488). When someone leaves, the account usually leaves with them. Fine for a few people; not for a whole team.

What is the biggest risk with relay stations? What you cannot see: whether the model answering is the one you paid for, whether the site will shut down, and whether there is a request record to check when something breaks. See the LLM API authenticity guide for how to test.

Can prices change suddenly? It depends on the route: read the provider's price-change notice rules before you sign or top up. Router One's notice period is on Router One for companies.

How is request data handled through a gateway? Go by each provider's published policy; Router One's is on data retention.

Can Router One issue invoices or take corporate bank transfers? See Router One for companies: invoices, payment methods, contracts, limits and refunds are all on that page.