Claude Code payments: subscription and API channels

separate “subscription” from “API”

• Searching for "how much does Claude Code cost per month" comes up empty — it is not a product sold separately and has no standalone subscription.
• Two channels: included in a Claude subscription (Pro / Max / Team / Enterprise), or pay-as-you-go API (the Console or a cloud provider).
• The official docs state: the free tier does not include Claude Code — you need one of the paid accounts above.
• The dividing line isn't price, it's the way you work: a person at the terminal uses the subscription; machines running on their own use the API.
• The two problems in China must be kept apart: payment can be solved; service-region support is not something a card can solve.

Claude Code has no price of its own

The question "how much does a Claude Code subscription cost" presupposes a product that doesn't exist. Claude Code is Anthropic's official client tool — it has no separate subscription and no separate price. Its money comes from two places:

• Included in a Claude subscription: the quota of plans like Pro / Max can be used directly by Claude Code, with no extra charge;
• Pay-as-you-go API: log in with an API key and costs accrue by token to your Console organization (or cloud account).

The two channels measure usage differently — don't take one channel's usage and convert it into the other's price.

On cost, the official docs add one key mechanism note: Claude Code is billed by API token consumption; the session cost estimate shown to subscription users is not a billing basis — what matters is the quota bar and the usage details. Subscriptions have usage windows and quotas; API billing follows token consumption.

There is also a route that doesn't go through a personal subscription: connect Claude Code to a cloud provider (such as Amazon Bedrock, Google Cloud, Microsoft Foundry) and bill by token to the cloud account. This route mainly serves teams with cloud contracts.

How to confirm in the terminal which channel you're on

The first time you run claude is the moment of choosing a channel. The official docs give two paths:

How to check which path you're on now: use the command from the official docs to look at the state — /usage shows quota and usage details (session cost estimates, per-model usage, cache hits, the subscription quota bar); use /login and /logout to switch accounts or re-authenticate.

The quota under a subscription is shared

On the subscription path there's an easily overlooked fact: Claude Code doesn't eat its own quota. The official docs state that subscription usage resets on a rolling five-hour window plus a weekly window and shares the same quota pool as the Claude web app.

• Consequence one: running hard in the terminal makes the web side hit its limit sooner — the two entrances affect each other;
• Consequence two: when judging "should I upgrade", looking only at web usage underestimates what's really being consumed — count terminal sessions too (for picking a tier, see Claude subscriptions: Pro, Max and billing).

When you hit the limit, the official prompts come in three kinds, with different meanings and responses:

Anthropic may change usage windows and reset timing. Follow the prompts and current usage details.

API channel: the cost structure decides how you use it

The two channels have completely different cost shapes: a subscription is a fixed monthly cap; the API is linear in usage with no cap. So on the API, the thing to manage isn't "which tier to buy" but "how much this session will burn".

The official docs give an order-of-magnitude reference (the average for enterprise rollouts): about $13 / developer / active day, $150–250 / developer / month, and 90% of users spend under $30 a day; there's also a small amount of background consumption even when idle (conversation summaries, prompt suggestions, etc.), which per the official wording is usually under $0.04 per session. These numbers are only for judging the magnitude, not your bill.

Three mechanisms that genuinely cut cost (all per the official docs):

• Model choice: Sonnet is enough for most coding tasks; leave Opus for complex architecture and multi-step reasoning. Switch mid-session with /model, or change the default in the config; subagents can be assigned a smaller model;
• Prompt caching: turned on automatically by Anthropic; the system prompt and repeatedly-read codebase context are billed at a lower unit rate when they hit the cache. Whether you hit it is visible in /usage;
• Context management: use /clear when switching tasks and /compact to compress; a long session left uncleaned re-computes the entire context on every turn.

Two more small items, also per the official wording: /effort and the environment variable MAX_THINKING_TOKENS can cap the extended-thinking budget; and using ready-made CLI tools (like gh, aws) instead of hanging a pile of MCP servers on reduces the context carried on every turn. Also, choosing data residency carries a rate factor of about 1.1x.

The two things that actually block users in China

Back to the headline question: how do you use Claude Code in China? Most tutorials answer with installation and network, but what actually decides whether you can keep using it long-term are two other things — and they must be looked at separately: one can be solved, one can't.

The first: payment — this layer can be solved

Claude Code's payment is just a Claude subscription or Console credit, so the gates you pass are exactly the same as subscribing to Claude: a BIN that supports AI-subscription scenarios, the card / billing address / payment network all aligned, and a buffer left on the card. The full triage for this layer is in ChatGPT payment declines: two-stage troubleshooting and Ten virtual card decline reasons and troubleshooting checks.

Subscribe first or API first: buying API credit directly with a freshly opened card gets blocked noticeably more often than running one monthly charge first. The path our support team has repeatedly verified is — first complete one subscription with the same card, build one successful transaction record, and switch to the API a day or two later. This isn't an official statement; it's an experiential path: in this scenario the card's transaction history is the cheapest trust credential there is.

The second: service-region support — no card can solve this one

Anthropic's terms of service and supported-countries list do not include mainland China (per the official supported countries page); the official docs also list "supported countries" as one of the running requirements. In other words: being able to install and run it and being "allowed to use it" are two different things — the former is a technical problem, the latter a terms-of-service problem.

Our position on this is clear: this site only handles the payment and subscription layer, and neither provides nor recommends any way to bypass regional restrictions, including any form of network workaround. Whether to use it in your region is yours to confirm against the terms of service and local laws.

Putting it back into cost terms makes it clearer: failing once on the card problem costs a few dollars and a little time; failing once on the region-compliance problem costs the account and the money already paid. They are not the same magnitude — Payment preparation does not replace regional eligibility checks.

Getting the money ready: this site's rate terms

If you've decided to go this route, the money to prepare on the payment side, per this site's current terms, is as follows:

The top-up advice follows the site-wide terms: leave a buffer above the list price for monthly plans — ≥$25 on the card for a Pro plan, and ≥$210 for high tiers like Max 20x. Don't top up to exactly the list price — that's a common failure scenario.

FAQ

Further reading

Claude Code's payment structure and channel choice — it doesn't repeat what other pages have already written: which subscription tier to buy, see Claude subscriptions: Pro, Max and billing; screen-by-screen subscription steps, see Claude subscription illustrated tutorial; the layered triage for declined payments, see ChatGPT payment declines: two-stage troubleshooting; bound successfully but the charge failed, see Card linked but payment failed: Claude API troubleshooting; after a failure, check the records first, see ChatGPT payment declines: two-stage troubleshooting; the ten root causes of declines, see Ten virtual card decline reasons and troubleshooting checks.

Concepts and the deposit pipeline are in U cards: definition and differences from virtual credit cards; card types and available regions, see all card types overview; rates, top-up guidance and withdrawal terms defer to the pricing page; single-point short answers are in the FAQ.

Official sources: installation and sign-in paths per Claude Code official setup docs (external source, for verification only, not an endorsement); usage and cost mechanisms (including /usage, prompt caching, model choice) per official cost docs (external source, for verification only, not an endorsement); features and subscription tiers, see the official product page (external source, for verification only, not an endorsement), the official docs overview (external source, for verification only, not an endorsement) and the official pricing page (external source, for verification only, not an endorsement); supported regions per the official supported countries page (external source, for verification only, not an endorsement). External links are for verification only and don't constitute an endorsement; commands, interfaces and rules change with versions — defer to the official docs of the current version.