Quick Answer
Claude is not officially available in mainland China through Anthropic's direct channels. However, Chinese users in 2026 have several workable methods: VPN + direct API (works but has latency and stability issues), mirror sites (risky, often unreliable), relay stations (common but many break prompt caching, making costs explode), and direct API gateways like TeamoRouter (the most stable and cost-effective approach). TeamoRouter provides a globally accessible API endpoint that works from China without a VPN, maintains >99% prompt cache hit rates, and supports all Claude models plus GPT, Gemini, and DeepSeek. This guide covers every method in detail.
The Claude-in-China Landscape: What Changed in 2026
The landscape for accessing Claude from mainland China has shifted significantly. Anthropic does not offer direct service in China -- claude.ai is blocked, and the official API (api.anthropic.com) is inaccessible without a VPN or proxy. Yet demand among Chinese developers, researchers, and businesses has only grown, driven by Claude's strong performance in coding, long-context reasoning, and Chinese-language tasks.
This has created a fragmented ecosystem of workarounds, each with distinct trade-offs in stability, cost, and safety. Below, we examine every available method in 2026, ranked from least to most recommended.
Four Methods at a Glance
| Method | VPN Required | Prompt Caching | Agent Compatible | Security | Best For |
|---|---|---|---|---|---|
| VPN + Direct | Yes | Yes (official) | Full | High | Occasional short sessions |
| Mirror Sites | No | No | Limited | Very Low | NOT RECOMMENDED |
| Relay Stations | No | Usually broken | Partial | Low-Medium | Casual chat (not agents) |
| TeamoRouter Gateway | No | Yes (>99%) | Full (100%) | High | Daily dev, agents, production |
Method 1: VPN + Direct Access
How It Works
- Subscribe to a reliable VPN with servers in Anthropic-supported regions.
- Connect to a supported-region server.
- Access claude.ai in your browser, or configure your API client to use
api.anthropic.com.
Pros
- Full official experience with no intermediaries
- Data goes directly to Anthropic's servers (through the VPN tunnel)
- Official prompt caching works
Cons
- VPN required at all times -- if it drops mid-session, Claude becomes inaccessible
- Latency penalty: China -> VPN server -> Anthropic API adds delay
- VPN blocks: Chinese network authorities periodically block VPN protocols
- Cost: Good VPNs add $5-$15/month on top of API or Pro costs
- Account risk: Anthropic may flag accounts consistently connecting from VPN IPs
When to use: Acceptable for occasional short Claude Web sessions. Not suitable for daily development or long Claude Code sessions.
Method 2: Mirror Sites (High Risk, Not Recommended)
Mirror sites are unauthorized third-party websites that replicate the Claude Web interface, routing requests to Claude's API behind the scenes. They are often free or suspiciously cheap.
Why to Avoid
- Severe security risk: Every message passes through the operator's server. They can read, log, and store all your conversations -- including code, business plans, and accidentally pasted API keys.
- Unreliable uptime: Mirror sites come and go. When the operator's keys get revoked, the site disappears overnight.
- No prompt caching: Operators rotate through accounts to avoid rate limits, breaking caching completely. For Claude Code users, this means 5-10x the token consumption.
- Model downgrades: Some mirrors silently substitute cheaper models while claiming to serve premium ones.
- Legal gray area: Using unauthorized mirrors violates Anthropic's terms of service.
Verdict: Avoid mirror sites for anything beyond casual experimentation. For professional work, the security risk alone disqualifies this method.
Method 3: Relay Stations (Common but Caveat Emptor)
Relay stations (中转站) are commercial API proxy services that accept your requests and forward them to Claude's API, handling cross-border routing on your behalf.
How They Work
- Sign up for a relay service and get an API key and base URL.
- Configure your Claude client to point at the relay's endpoint.
- The relay forwards your request to Anthropic's API and returns the response.
The Quality Spectrum
- Enterprise-grade relays: Operate dedicated upstream accounts with consistent routing. Essentially API gateways (see Method 4).
- Mid-tier relays: Mix of dedicated and shared accounts. Caching may work intermittently.
- Budget relays: Rotate through pools of shared/low-cost accounts. Caching is effectively non-existent. These are the heavily advertised "90% off" services.
The Hidden Cost: Broken Prompt Caching
This is the single most important factor. Claude Code, Codex, and other agent tools resend large contexts on every turn. With working caching, cached tokens cost only 10% of full price. Without caching, you pay full price for the entire context on every turn. A session that should cost $2 can easily become $20-$50.
How to Evaluate a Relay
Before committing, test: send two identical requests in quick succession. If the second request's token usage is significantly lower (cached tokens appear separately), caching works. If not, walk away -- your real costs will be far higher than the sticker price.
Method 4: TeamoRouter Direct Gateway (Most Recommended)
A direct API gateway is the professional-grade solution. Unlike relay stations that resell API access, TeamoRouter operates its own high-quality upstream channels with dedicated infrastructure and enterprise-grade reliability.
How It Works
- Sign up at teamorouter.com and get your API key.
- Point your Claude client at TeamoRouter's globally accessible endpoint:
https://api.teamorouter.com/anthropic. - No VPN needed. TeamoRouter handles cross-border routing on optimized network paths.
- All Claude models available, plus GPT, Gemini, DeepSeek through the same key.
Why TeamoRouter Is the Best Option for China
1. Prompt caching that actually works (>99% hit rate)
TeamoRouter achieves an extremely high cache hit rate through premium upstream channels and low-level optimizations. Combined with floating rates as low as 10-20% of official pricing, your effective per-token cost for agent workloads is a fraction of what budget relays charge.
2. No VPN, no fuss
The API endpoint is globally accessible. Configure once and it just works. No VPN to maintain, no tunnel to drop mid-session, no protocol blocks to worry about.
3. 100% agent-protocol compatible
Claude Code relies on specific API features (tool calling, streaming, extended thinking, beta headers). TeamoRouter maintains full protocol compatibility, regularly tested against Claude Code, Codex, and Gemini CLI.
4. Enterprise-grade reliability (SLA 99.6%)
Up to 5,000 QPM, tens of millions of TPM. Time-to-first-token is close to direct-Anthropic levels. When you run an agent overnight, you need absolute stability.
5. Unified access to all models
Your key works across Claude, GPT-4o, Gemini, DeepSeek, and more. One key, one dashboard, one bill. See TeamoRouter's AI API gateway for details.
6. Chinese-friendly payment
Accepts Alipay and WeChat Pay. Dashboard available in Chinese. Support team understands Chinese users' challenges.
Quick Setup
export ANTHROPIC_API_KEY="tr-your-teamorouter-key"
export ANTHROPIC_BASE_URL="https://api.teamorouter.com/anthropic"
# Persist
echo 'export ANTHROPIC_API_KEY="tr-your-key"' >> ~/.zshrc
echo 'export ANTHROPIC_BASE_URL="https://api.teamorouter.com/anthropic"' >> ~/.zshrc
source ~/.zshrc
# Install and run
npm install -g @anthropic-ai/claude-code
claude
Common Pitfalls for Chinese Claude Users
Pitfall 1: Focusing Only on Sticker Price
Choosing a relay solely because it advertises "90% off" is the most common mistake. A relay charging 1 cent per 1K tokens with zero cache hit rate costs more than a gateway charging 5 cents with 90% cache hit rate. Always test caching before committing.
Pitfall 2: Using Claude Code Without Understanding Caching
Many developers install Claude Code, connect to the cheapest relay, and get shocked by a $200 bill after a weekend of heavy use. Claude Code sends large contexts on every call -- without caching, a single session can consume millions of tokens.
Pitfall 3: Mixing VPN and API Gateway
Using a VPN for claude.ai and a relay for Claude Code simultaneously creates configuration conflicts. Pick one consistent access method.
Pitfall 4: Ignoring the TeamoRouter Client
The TeamoRouter Client auto-configures your environment for optimal routing from China -- proxy settings, environment variables, connectivity diagnostics. Many users struggle with manual configuration when the client solves their issues in one click.
Method Comparison Table
| Criteria | VPN + Direct | Mirror Site | Budget Relay | TeamoRouter Gateway |
|---|---|---|---|---|
| VPN required | Yes | No | No | No |
| Setup complexity | Medium | Low | Low-Medium | Low |
| Prompt caching | Yes (official) | No | Usually broken | Yes (>99%) |
| Agent compatibility | Full | Limited | Partial | Full (100%) |
| Uptime reliability | VPN-dependent | Very low | Variable | 99.6% SLA |
| Security | High | Very low | Low-Medium | High |
| Latency from China | VPN overhead | Low (unreliable) | Variable | Low (optimized) |
| Model selection | Claude only | Limited | Claude-focused | Claude + GPT + Gemini + more |
| Payment | VPN fee + API cost | Often free | Alipay/WeChat | Alipay/WeChat |
| Best for | Occasional short sessions | NOT RECOMMENDED | Casual chat | Daily dev, agents, production |
FAQ
Can I use claude.ai directly in China without a VPN?
No. claude.ai is blocked in mainland China. Direct access results in a connection timeout. A VPN is required for claude.ai access. For API access (Claude Code, Claude CLI), a gateway like TeamoRouter works without a VPN.
Is using Claude through a relay station legal?
Using a relay or API gateway is not explicitly illegal in China but operates in a regulatory gray area. Anthropic's terms prohibit unauthorized resale of API access. Gateways with proper upstream agreements are on firmer ground. Your primary risk is service disruption, not legal liability -- but avoid mirror sites using stolen credentials.
Why does my Claude Code bill feel so high from China?
The most likely cause is broken prompt caching. If your provider rotates accounts, caching does not work. Since Claude Code resends long contexts every turn, token consumption can be 5-10x higher than expected. Switch to a provider with verified caching and costs should drop dramatically.
What is the latency like from China to TeamoRouter?
TeamoRouter uses optimized network routes with edge servers in nearby regions. Typical round-trip latency is 100-300ms -- comparable to a good VPN connection. For Claude Code sessions, model processing time dominates, so the per-turn latency is barely noticeable.
Can I use TeamoRouter on my phone in China?
Yes. Use third-party chat clients like ChatBox, LobeChat, or NextChat configured with your TeamoRouter key. These are available on iOS and Android and provide a Claude-backed chat interface.
Will Claude ever become officially available in China?
As of 2026, Anthropic has not announced plans for official China availability. The regulatory environment for foreign AI services is complex. For the foreseeable future, API gateways will remain the primary access method. Gateway technology has matured enough that the experience is nearly indistinguishable from direct access.
What happens if TeamoRouter goes down during a Claude Code session?
Claude Code handles interruptions gracefully -- it retries with exponential backoff, and your session state is preserved locally. TeamoRouter's 99.6% SLA and redundant upstream channels make such interruptions rare.
Can I switch from my current relay to TeamoRouter without losing my setup?
Yes. Change two environment variables: ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY. Claude Code itself does not change. Your project settings, CLAUDE.md, session history, and MCP configs all remain intact.