Claude API in China: Access, Payment & Setup (2026)

Claude remains a favourite for coding and long-document work, but the official channel puts three hard requirements in front of most developers in China: an overseas phone number, an overseas payment method, and passing risk review after signup. Here are the realistic paths, with runnable code.

Path 1: Official direct (teams with an overseas entity)#

If you have an overseas company and card, console.anthropic.com is the cleanest route. Three caveats:

• Prepaid, $5 minimum
• Strict risk controls — changing IPs can trigger reviews
• Rate limits scale with deposits; new accounts are tightly capped

Without an overseas entity, don't force it. Use path two.

Path 2: Official relay (most people)#

An "official relay" buys official API capacity wholesale and resells it. How to tell a real one:

1. Behaviour matches the official model: thinking, tool use and image input all work
2. Billing mirrors the official structure: per-token, input and output priced separately
3. Sane pricing: typically 40–60% below list. Absurd discounts mean reverse-engineered channels

zaimod's Claude pricing at the default group (November 2026, USD per 1M tokens):

Top-ups accept local payment methods and credits don't expire — the main convenience over official prepaid billing.

Two ways to connect#

A. OpenAI-compatible endpoint (smallest change)#

from openai import OpenAI

client = OpenAI(
    api_key="sk-your-zaimod-key",
    base_url="https://api.zaimod.com/v1",
)

resp = client.chat.completions.create(
    model="claude-sonnet-5",
    messages=[{"role": "user", "content": "Review this code for race conditions"}],
    stream=True,
)
for chunk in resp:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

B. Anthropic-native endpoint (full feature set)#

Thinking blocks and native tool use stay intact via /v1/messages:

import anthropic

client = anthropic.Anthropic(
    api_key="sk-your-zaimod-key",
    base_url="https://api.zaimod.com",
)

msg = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=4096,
    messages=[{"role": "user", "content": "Write an async downloader with retries"}],
)
print(msg.content[0].text)

Reliability playbook#

1. Primary + fallback: claude-sonnet-5 as primary, gpt-5.6-terra or gemini-3-pro-preview as backup
2. Retry with backoff: exponential (1s → 2s → 4s) on 429/5xx; never hot-loop
3. Reconcile daily: a token-usage spike usually means a missing loop terminator

Common pitfalls#

• Missing max_tokens: Claude requires it explicitly on the native endpoint; the OpenAI-compatible layer substitutes a default but don't rely on it
• System prompt placement: in the Anthropic protocol, system is a top-level parameter, not a message role
• Long-context cost: stuffing a whole repo into 200K tokens costs dollars per call. Retrieve first, then inject context

zaimod carries the full Claude lineup — signup takes two minutes →

Further reading#

• Claude native /v1/messages docs
• How to Choose an OpenAI API Gateway
• Live Claude pricing