Using Claude Code from China — Install, Subscription, and API Payment Guide

Claude Code installs and runs fine from China; the step that actually stops people is payment. Installing it is one npm command, and pricing splits into two completely different paths: a subscription allowance and API billing. Payment failures have to be read in two layers. First, check whether the card BIN range itself supports AI subscription scenarios — for some ranges the issuer policy explicitly does not, and taking one of those to an AI subscription fails no matter how much you troubleshoot; switching ranges is the fix.Once the BIN range is right, the card itself tests clean, and the failures that remain land on the user side and the account side (limit, card status, balance, an account that is too new, an unstable IP). This article runs in four parts — what it is, how to install it, how to pick a pricing path, how to get the payment through — andthe payment section, the part almost every guide skips and the only part where we hold first-hand data as a card issuer, is written in the most detail.

Assumed reader: a developer who has no virtual card yet and holds only domestic Chinese bank cards. If you already have a card and only want the error fixed, jump straight to what to do after a declined payment.

1. What Claude Code Is — and How It Relates to Claude on the Web

Claude Code is Anthropic’s official command-line coding assistant: it installs on your own machine, runs in your terminal, and reads and writes files in your local repository, runs commands, runs tests, reads the errors and revises. It is not a chat box in a browser; it is an agent with hands.

Three ways it relates to Claude on the web

• Same account: the account you log in with on the web is the account Claude Code uses. You do not need a second one.
• Same models: the difference is in how you use them — on the web you copy and paste code, in Claude Code it edits the files directly.
• Billing can be shared or separate: this is the crux of the article. A subscription allowance (Pro / Max) can be spent by Claude Code directly, or you can skip the subscription entirely and pay per use through the API. The two paths differ in cost structure and in how hard they are to pay for. See section 3.

Which one to reach for: just asking questions or thinking something through, the web version is enough;when you want it to actually change code, run commands, and carry one job to completion, use Claude Code.

2. Installing Claude Code from China — Environment, Commands, Network

Short version: the install itself is easy, one npm command. From China, all of the friction is network. And the network does not only affect installation — it directly decides whether adding your card later goes through (see section 4).

Requirements

• Node.js 18 or newer (needed for the npm install path; an older runtime errors out during installation)
• A Unix-like shell: macOS and Linux work natively; on Windows, run it inside WSL
• A stable connection out of the country: it talks to Anthropic endpoints continuously, and a dropout kills the session

Install and first run

# 1. Global install (slow from China? route just this one command through a mirror)
npm install -g @anthropic-ai/claude-code --registry=https://registry.npmmirror.com

# 2. cd into your project before starting it - the current directory is the workspace
cd /path/to/your/project
claude

The first time you run claude, it asks how you want to sign in: with a Claude subscription account (Pro / Max, through a browser authorization), or with an Anthropic Console account (API billing). That prompt is exactly the fork section 3 wants you to settle first — picking wrong is not permanent, it just costs you a round of setup.

Three things to do about the network in China

1. Use a mirror for the install only: the --registry flag at the end of the command above fixes the slow download, and it applies to that single install — it does not rewrite your global npm registry. Mirrors lag behind, so once it is installed you can check the version against the official registry.
2. Set a proxy for the terminal when you run it: a GUI proxy client does not automatically cover terminal processes. You have to exportHTTPS_PROXY / HTTP_PROXY explicitly in your shell (with whatever port your client actually listens on). This is the single most common reason the website opens in a browser while the terminal cannot connect.
3. Pin the exit node: turn off auto-switching and load balancing. Hopping between IPs barely matters for day-to-day use, but it is fatal when you add a card — the specific requirements are the 5 prerequisites in section 4.

A few commands you will use once it is running

• /login / /logout — switch accounts or sign in again
• /status — see who you are signed in as and which billing path you are on
• /cost — usage stats for the current session (which fields it shows depends on your own build)
• /model — switch models; /clear — wipe the context and start over (do this first when a long session gets slow and expensive)

3. Claude Code Pricing: Subscription Allowance vs API Billing

Conclusion first: the large majority of individual developers should take the subscription, not the API. A subscription buys a usage window for a fixed monthly fee, so cost is predictable; the API bills per token, which feels great while you write and runs up a bill just as fast. More to the point — on the payment side, the two paths differ by an order of magnitude in difficulty (see section 5).

What the subscription side costs

Individual plans come as Pro and Max. List prices are $20 per month for Pro and two Max tiers at $100 and $200 per month (annual billing is discounted; go by whatever the official pricing page lists at the time). What separates the three tiers is the size of the usage window, not the feature set.

A list price is not necessarily the amount that hits the card: US sales tax rates differ by state, so two people can legitimately be charged a dollar or two apart. The Claude subscription charges we see in our own ledger are mostly round figures like $20.00 and $200.00, with no visible tax spread (our ledger basis, as of 2026-07-17). The takeaway is not that you will be overcharged, and it is certainly not that you should go fiddle with addresses — it is do not fund the card right at the list price. Section 4 gives the exact amounts to budget.

How to choose between Pro and Max, and how the allowance is actually counted, are not repeated here. See choosing a Claude subscription (Pro / Max 5x / Max 20x) and how Claude usage limits work and what to do when you run out.

What the API side costs

API pricing is per million tokens (MTok), counted separately for input and output. As checked on 2026-09-20, the three tiers most people write code against sit roughly at a high tier around $5 / $25 per MTok (input / output),a mid tier around $2 / $10, and a light tier around $1 / $5; above those sits a more expensive flagship tier. Model names and unit prices both change, so go by the official API pricing page and do not copy numbers out of any tutorial — including this one.

Two mechanisms genuinely lower the bill and are worth knowing: on a prompt-cache hit, the re-read portion bills at roughly a tenth (writing to the cache costs slightly more), and asynchronous batch jobs run at roughly half price. If your pattern is feeding the same large repository context over and over, caching makes an enormous difference.

A way to decide without guessing

Do not pick on instinct. Match how you actually work against the list below; this step needs no arithmetic:

• You use it nearly every day, hours at a time → subscription, capped monthly fee, no surprise bill
• Four days a week you barely touch it, then three days you burn hard → API fits better, pay per use instead of paying for idle time
• You need it inside CI, in scripts, or split across a team by project → API only, a subscription is for a person, not a pipeline
• You keep hitting the window cap and the wait for a reset is costing you work → first consider moving up a subscription tier, then consider the API (the allowance article linked above explains the counting)

There is exactly one reliable way to settle the arithmetic: sign in with an API key for a week, then check the actual bill in the Console, and compare it to $20 / $100 / $200 — equivalent spend clearly below the monthly fee means you are a light user and per-use is cheaper;above it, the subscription wins, and it caps your downside. The reverse does not hold: a subscription allowance and API billing are two different measurement systems, sodo not take the usage readout from one and convert it into a price for the other — the number that comes out cannot support a decision.

4. Getting the Claude Code Payment Through — the Whole Card-Adding Procedure

Conclusion: pick the right BIN range first, troubleshoot second. This section runs in two layers — layer one is whether the BIN range is right,layer two is the environment and the account. Reverse the order and everything you do afterwards is wasted.

Layer two: only once the BIN range is right do user-side and account-side causes come into play. On a BIN range that supports AI subscription scenarios, the card itself tests clean, so a failure at that point comes down to the limit, the card status, the balance, an account that is too new, or an unstable IP. Everything below — including the failure-cause split — describes that second layer. But before any of it, one more thing has to be separated out: card-binding verification is not a charge, and the two do not trip up the same people at all.

Those numbers carry one more premise: they count authorizations that actually reached the card side. A separate class of failure never gets that far — the merchant’s own pre-authorization risk checks block it the moment you submit, andthe card transaction history shows nothing at all. That class has nothing to do with the card, swapping cards usually changes nothing, and what needs fixing is account age, IP stability, and whether the billing address is in the same country as your IP. Section 6 shows how to tell at a glance which class you hit.

4.1 Get the payment instrument ready first

Domestic Chinese bank cards — including the ones banks badge for overseas spending — essentially do not work in this scenario, and this article does not cover that route. What does work is a US-BIN virtual credit card plus the matching US billing address. If this is entirely new to you, read what a virtual credit card is first, then come back.

When you pick a BIN range, only one thing matters: does it support AI subscription scenarios. Issuer policies differ from range to range, the supported scenarios are written into the BIN range notes, andwhat the card-issuing page shows is what counts — do not decide from someone’s screenshot or a rumor that a given range works. For the full list of ranges see US Visa BIN ranges; for AI subscriptions the recommended product is the shared-limit card.

4.2 The 5 prerequisites you have to meet before adding the card

These five came out of our support team verifying them over and over on the front line, and they describe layer two —the premise being that the BIN range of the card in your hand is already confirmed to support AI subscription scenarios. If that first layer is not settled, scoring full marks on these five changes nothing.Miss any one of them and, even if the payment slips through, you are likely to be asked for KYC documents later or to lose the order outright.

1. ① The account is at least a week old and was used normally during that week

   Attaching a card to an account less than a week old — especially one just registered through a proxy — usually fails, and even when it passes it tends to trigger a KYC request. The right order is: register, chat normally for a few days, add the card after a week. Plenty of people do it backwards, rushing to pay on signup day and then blaming the card.

2. ② Keep the IP stable for 24 to 48 hours before you add the card

   Never jump across countries or regions. If you switched nodes in the past couple of days — even just Los Angeles to New York —pin it down, let it sit for 24 hours, and add the card after the observation window passes. Auto-switching and load-balancing settings in your client must be turned off beforehand.

3. ③ The IP must be yours alone; datacenter versus residential matters less than you think

   Plenty of guides insist you must use a residential IP. What we actually observe is thatthe gap between datacenter and residential IPs is small, and the fatal property is sharing — public proxy nodes and exit IPs shared by many people were very likely blacklisted long ago, and no amount of cleanliness on your end fixes that. Residential IPs are slightly better than datacenter ones, but if you have to choose, choose exclusive.

4. ④ Card, IP, and billing address in the same country

   Card country = the country of the proxy IP you pay from = the country of the billing address you enter at checkout. Ideally go one step further:IP state = billing address state. When this one is off, the failure usually happens the instant you submit, and nothing is recorded on the card side. How to fill in the billing address is covered in 4.3.

5. ⑤ Leave headroom in both the card balance and the limit; do not fund at exactly the list price

   A $20 list price does not necessarily mean $20 comes off the card: state sales tax rates differ, the actual charge can land a dollar or two off, and a retry after a failure needs room too. Funding exactly to the list price is where failures cluster — look again at the attribution above (on BIN ranges that support AI scenarios): about 80% of charge failures were an insufficient per-transaction limit or available limit. For a $20 monthly plan, keep the available card balance at $25 or more (the Claude ledger shows no tax spread and no retry-charge pattern, so $25 is enough; you do not need the $26 we suggest on the ChatGPT side); for a $200 annual plan or Max 20x, budget $210 or more.

4.3 How to fill in the billing address

Some details of the form itself, all of them frequent trip-ups:

• Prefer the set assigned with the card, copied character for character into the checkout page. Do not type it from memory.
• Line 1 is the street only. Do not pile the city and state into it.
• Leave line 2 (Address line 2) empty — no apartment number, no “N/A”, no space character. This is a high-frequency, entirely avoidable failure.
• City, state, and ZIP have to agree with each other (a ZIP that does not belong to the state is a problem).
• ⛔ Do not copy a “universal address” circulating online: thousands of people sharing one address is itself a risk signal.

4.4 The full checkout procedure (7 steps)

1. Confirm the BIN range supports AI subscription scenarios

   The BIN range notes on the card-issuing page state which scenarios are supported, and what is shown there is what counts. This step is first and it is not filler — wrong range, and the other six steps are wasted.

2. Check the account state

   At least a week old, with normal use during that week. If it is not, wait — this step cannot be skipped.

3. Pin the network and let it sit for 24 to 48 hours

   Pick one exclusive node, pin it, turn off auto-switching. Before you submit, check yourself once with an IP lookup site: country, state, and IP type all as expected.

4. Prepare the card with headroom in both balance and limit

   For the $20 monthly plan budget $25 or more, for an annual plan or Max 20x budget $210 or more. Do not fund exactly to the list price. The limit is the total pool you set when the card is issued (the minimum issuance limit is whatever the card-issuing page shows), renewals draw from that same pool, so if you intend to subscribe long term, size it against total spend.

5. Subscribe on the web, not through mobile in-app purchase

   In-app purchase goes through the app store, usually costs more, and runs on a payment rail separate from your card.

6. Enter the card details and the billing address

   Card number, expiry, CVC — always use the copy button, never retype them. In real failure records, a mistyped CVC or a reversed expiry date is a frequent cause. Fill in the billing address per 4.3: copy the set from the card detail page, street only on line 1, line 2 empty.

7. Submit exactly once

   On success, note the amount actually charged to the card and compare it to the list price (the dollar or two of difference is usually state sales tax).If it fails, stop immediately — do not re-enter, do not click again. Go to section 6.

5. API / Credit Top-Ups: Why They Are So Much Harder Than a Monthly Plan

Conclusion: the risk controls on a Credit or API top-up are nothing like those on a monthly subscription, and the bar for the account, the card, and the environment is a notch higher everywhere. This comes out of front-line support experience, and it is a stretch most Claude Code tutorials omit — the people writing them mostly bought a monthly plan and never walked the API path.

The classic failure case

A freshly issued card with no prior successful transaction, taken straight to buying API credit, is usually declined. The same card in the same environment will often clear a monthly subscription on the first try. A lot of people conclude at that point that the card is bad and go get another one — and the next few cards behave the same way.

The path customers worked out themselves, confirmed over and over

1. Use the card for a monthly subscription first (Pro is fine), so it generates one successful transaction record.
2. Wait a day or two rather than continuing the same day.
3. Then buy Credit or API balance, and the proportion that goes through rises noticeably.

We are not going to explain how the upstream scores its risk model; we only report that this path reproduces in practice:a card’s transaction history is the cheapest and most effective trust signal available in this scenario. One related suggestion: if the Console dashboard offers automatic reloading (auto-reload),do the first two top-ups by hand — every charge is an independent risk decision, and before the card has a history, letting it run unattended means a failure happens without you knowing which attempt it was.

This is also what the last row of the section 3 table means: for someone with no card who wants to go straight to the API, the path is “subscribe to Pro for one month to season the card, then move to the API”, not going straight for Credit on day one.

6. What to Do After a Declined Payment — Rapid Retrying Is the Costliest Move

Cooldown strategy

The exact number of hours is experiential and varies by merchant and by timing, but the direction — the more you retry, the worse it gets — is consistent. A note on cost while we are here: on a dedicated-limit card, each failure also carries a separate decline fee (currently $0.60 per attempt; shared-limit cards do not charge it, and the rates are whatever the card-issuing page and the pricing page show). Clicking five times in a row does not only lock you out; it costs real money.

Troubleshooting order: BIN range first, then who declined it

Step one: go back to the card-issuing page and confirm this card sits on a BIN range that supports AI subscription scenarios. It takes under thirty seconds and it decides whether everything after it means anything: if the issuer policy behind that range does not support AI subscriptions in the first place, the failure is an unsupported BIN range, moving to a range that supports AI scenarios solves it, and no amount of troubleshooting, no number of replacement cards on the same range, and no amount of IP hygiene will help.

Step two: once the BIN range checks out, open the card transaction history and see whether this attempt left a record. This is the fork that saves the most time:

• A record with an explicit reason (insufficient per-transaction limit, insufficient balance, wrong expiry or CVC) — fix what the message says. The attribution in section 4 describes exactly this class: on BIN ranges that support AI scenarios, about 80% insufficient per-transaction limit and about 20% mistyped expiry or CVC (our ledger basis, per-authorization attribution for ANTHROPIC / CLAUDE merchants, as of 2026-07-17; distributions differ between merchant families, so do not carry this set over to another one).
• No record at all — the decline happened in the merchant’s pre-authorization risk checks and has nothing to do with the card. At that point switching cards is pure waste; what to examine is account age, IP stability, and whether the billing address is in the same country as the payment IP (same state is better).

For the complete list sorted by cause, see the top 10 reasons a virtual card gets declined; for a step-by-step troubleshooting flow, see troubleshooting a failed ChatGPT subscription paid from a Chinese card — that one is written about ChatGPT, but the reasoning and the BIN-range-first order apply to Claude just as well.

7. FAQ

Can Claude Code be used from China?

Technically it installs and runs, and section 2 of this article is the full procedure. But the other half has to be said plainly: Anthropic’s terms of service and supported regions do not include mainland China, and signing in directly from a mainland network carries a real risk of the account being restricted. That is not something a card solves. For where the risk boundary sits, see the real reasons Claude / ChatGPT accounts get banned.

Do I have to subscribe to Pro to use Claude Code?

Not necessarily. Claude Code supports two billing paths — a subscription allowance (Pro / Max) and per-use API billing — and you choose between them on first run. But if you do not yet hold a card that can pay,in practice we suggest one month of Pro first; the reason is in section 5, since a brand-new card buying API credit outright is usually declined.

Can a domestic Chinese credit card pay for Claude Code?

Essentially not, and it is not worth spending time on repeated attempts (which also lock you out after consecutive failures, see section 6). The workable route is a US-BIN virtual card plus a US billing address: first confirm the BIN range supports AI subscription scenarios (the BIN range notes on the card-issuing page state which scenarios are supported, and that is what counts), then line up the 5 prerequisites in section 4 one by one.

Why do some cards never manage to subscribe to Claude?

Before you start rebuilding your environment, look at the BIN range.For some BIN ranges, the issuer policy explicitly does not support AI subscriptions, and taking such a range to Claude or ChatGPT fails because of an unsupported BIN range — that is a policy matter, not a probability one, and aligning account age, IP, and billing address perfectly will not change it. Moving to a BIN range that supports AI subscription scenarios is the fix.Conversely, if the range does support AI scenarios and the card tests clean, then the cause of the failure lands on the user side and the account side (limit, card status, balance, an account that is too new, an unstable IP), so work through the 5 prerequisites in section 4 and the fork in section 6. One more reminder: when this site says pass rate, it means the issuer scenario-support rate, not an authorization approval rate. Do not read the two as the same thing.

Do Claude Code and Claude on the web share the same allowance?

On the subscription path, yes — it is one usage allowance under one account, so heavy Claude Code sessions will also push the web version into its window cap sooner. For how the allowance is calculated and what to do when you hit the cap, see how Claude usage limits are calculated.

Can a subscription allowance feed a third-party command-line tool?

No. The OAuth credential a subscription issues is authorized for official Claude Code only, and a third-party tool getsThis credential is only authorized for use with Claude Code straight back. To use Claude inside another tool, take the API key route.

The card verified fine, so why did the charge still fail?

Because they are two separate gates. Verification is a $0 authorization that only proves the card details were entered correctly and the card is live — on our ledger basis (as of 2026-07-17) that gate has zero recorded failures, and what actually fails is the real charge afterwards. A charge still has to clear the card balance, the issued limit, and merchant risk controls.On a BIN range that supports AI subscription scenarios, the card itself tests clean, so the cause of a failure at that point lands on the user side and the account side: our per-authorization attribution of failed authorizations at ANTHROPIC / CLAUDE merchants (as of 2026-07-17) is about 80% insufficient per-transaction limit and about 20% mistyped expiry or CVC.Distributions differ between merchant families, so do not carry this set over to another one. The least troublesome prevention is simply not funding exactly to the list price. (If the BIN range does not support AI subscriptions in the first place, that is a different matter — see the previous answer.)

How much money do I need in total for one month of Pro?

Budgeting around $30 leaves plenty of room: $20 for the monthly fee, plus headroom (keep the available card balance at $25 or more; the Claude ledger shows no tax spread and no retry-charge pattern, so $25 is enough), plus the one-off card issuance fee, starting at $1 each and varying by BIN range; the shared-limit card recommended for AI subscriptions is $2, shown right on the card-issuing page.

The other fees, stated plainly:topping up your account carries a 0.5% fee, and USDT is usually credited within minutes (the minimum amount is whatever the top-up page shows);moving account balance onto a card (adding limit to it) carries a card top-up service fee starting at 2%, stepping down by calendar-month cumulative card top-up volume to a floor of 0.5%;pulling a card balance back to the account is free, but it requires 48 hours since that card’s last transaction, and some BIN ranges require a small balance to remain on the card, per the card detail page; returning account balance to your own USDT wallet is called a prepayment refund,a flat $2 per transaction, including payment processor fees — that is not a withdrawal, and the account is not closed. Both the rates and the configurable items are whatever the card-issuing page, the top-up page, and the pricing page show.

Can I just enter any US address as the billing address?

Not advisable. The billing address on the card detail page is exactly the set of details meant to be copied into the merchant checkout page, and using it as-is should be your default. The one thing that genuinely matters is this:the billing address has to be in the same country as the IP you pay from, ideally the same state.Whether you may substitute your own set, and whether doing so affects payment, varies by BIN range, so go by what the card detail page shows and says, rather than applying a blanket rule from some tutorial. When you are declined, try the address that came with the card first instead of swapping one in immediately. ⛔ And do not copy a “universal address” circulating online — thousands of people sharing one address is itself a risk signal.

I have no card at all. What is the shortest path?

The order is: register a Claude account and use it normally for a week → during that week, issue the US virtual card (pick a BIN range that supports AI subscription scenarios, per the card-issuing page) and fund the card balance to $25 or more → pin an exclusive node and let it sit for 24 hours → subscribe to Pro on the web and submit exactly once → once that works, decide whether to move to the API. Run the week of account age and the card issuance in parallel and the total wait is one week, no longer.

8. Related Reading

• ▸ How to choose a Claude subscription (Pro / Max 5x / Max 20x) — decides which tier Claude Code runs on
• ▸ How Claude usage limits work and what to do when you run out
• ▸ The complete Cursor subscription guide — the payment-side prerequisites are nearly identical
• ▸ What a ChatGPT subscription actually costs — if you want both, settle the arithmetic first
• ▸ Cross-border payment 101 — start here if BIN and billing address are new words to you