How to Use the Claude API from China (2026): Three Options Compared

This page compares the three, then shows the third-party route step by step with code we ran on November 1, 2026 against api.foo using a test key (the server answered 401 无效的令牌 — "invalid token" — which proves the URL, path and request format are right without spending money). For Claude prices, see Claude API pricing.

The three options at a glance

Whichever route you pick, your prompts leave your machine and go to that provider. Read its privacy terms before sending sensitive data, and follow the rules that apply where you operate.

Option 1: the official Claude Console

Sign up at platform.claude.com, add a card under Settings → Billing, buy credits, then create a key under Settings → API keys. Keys start with sk-ant- and are shown once. The full walkthrough, including spend limits and the 429 you get at the monthly cap, is in How to get a Claude API key.

This only works if you, or your company, are in a supported region with a card that the Console accepts.

Option 2: Claude on a cloud platform

Anthropic's models are also sold through AWS Bedrock, Google Vertex AI and Microsoft Foundry. You pay through the cloud bill and authenticate with cloud credentials instead of an Anthropic key. Anthropic's SDKs ship a client for each platform, so the request body stays the same while the client class and credentials change.

The catch is setup: you need a cloud account in a region where Claude is offered, model access must be enabled, and some features lag behind the direct API. It makes sense when your company already runs there.

Option 3: a third-party API service, step by step

A third-party API service (often called an API relay or gateway in China) gives you its own key and forwards your requests to the model. With the official SDK, only the address and the key change. Using api.foo as the example:

1. Sign up and top up. Create an account at api.foo and top up in Console → Wallet with Alipay, WeChat Pay, a card or USDT. Balances are in yuan; per-model prices are on the Models page.
2. Create a key. In Console → API Keys, click Create API Key and copy the key (it starts with sk-). Check that it can use Claude models.
3. Point your code at the service. Store the key in an environment variable — export APIFOO_API_KEY="sk-..." on macOS / Linux, or setx APIFOO_API_KEY "sk-..." on Windows (then open a new terminal) — and set the base URL.

With the anthropic SDK (pip install anthropic), stop the base URL at the domain — the SDK adds /v1/messages itself:

import os
import anthropic

client = anthropic.Anthropic(
    api_key=os.environ["APIFOO_API_KEY"],
    base_url="https://api.foo",
)
msg = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Introduce yourself in one sentence."}],
)
print(msg.content[0].text)

If your code already uses the OpenAI SDK, keep it and add /v1:

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["APIFOO_API_KEY"], base_url="https://api.foo/v1")
resp = client.chat.completions.create(
    model="claude-sonnet-5-5",
    messages=[{"role": "user", "content": "Introduce yourself in one sentence."}],
)
print(resp.choices[0].message.content)

Prefer the native anthropic SDK for anything that uses prompt caching, extended thinking, PDFs or images — those features are most complete on the Messages API.

To check the setup without spending anything, send one request with a deliberately wrong key. Both snippets above returned AuthenticationError ... 401 ... 无效的令牌 in our test: the request reached the right endpoint and was rejected only because of the key. A 404 or a connection error instead means the address is wrong.

Using Claude Code rather than your own code? See Claude Code with a custom API.

Four things to test before you commit to a provider

Claude has a few features that providers get wrong more often than others. Top up a small amount and check:

1. Prompt caching works. Add "cache_control": {"type": "ephemeral"} to a long request, send it twice, and look for a non-zero cache_read_input_tokens in the second response's usage. If it stays 0, long conversations will cost several times more.
2. Official model names work. claude-sonnet-5-5 and claude-opus-5-5 should be accepted as-is. Claude Code sends these names, so this matters if you use it.
3. Long prompts are billed the way Anthropic bills them. Haiku 5.5 costs five times more above 100k input tokens on the official price list; see whether the provider follows that rule or uses its own.
4. Thinking, tool use, images and PDFs work on the native /v1/messages endpoint.

For a general checklist (refund rules, who runs the service, uptime), see OpenAI API alternatives.

What to avoid

• Resold official keys. An sk-ant- key sold by an individual belongs to someone else's account. It can be shared, stolen or revoked at any moment, and your prompts pass through a stranger.
• Providers that hide who runs them. Look for terms of service, a privacy policy, a support contact and a refund rule before you top up.
• Large top-ups on day one. Start small and scale up once the four checks above pass.

Errors you may see

These messages come from New API, the open-source gateway that api.foo and many Chinese providers run on. More in LLM API errors explained.

Related

• Claude API pricing and cost calculator
• How to get a Claude API key
• Claude Code with a custom API or base URL
• What is api.foo?

Sources

Anthropic documentation, checked November 1, 2026: supported regions, Claude in Microsoft Foundry, prompt caching, pricing. OpenRouter FAQ for its payment methods (checked November 1, 2026).