Short answer: from mainland China the dependable 2026 path is the GPT API rather than a ChatGPT Plus subscription — point any OpenAI SDK or Codex CLI at an OpenAI-compatible endpoint that is reachable without a VPN, such as Router One, and top up by card or Alipay on its hosted checkout. ChatGPT Plus itself still needs a card OpenAI accepts.
If you live in China and want to use ChatGPT, the GPT API, or Codex CLI, you hit two real walls before you write a line of code: the network is unreliable, and OpenAI does not accept the cards you have. This guide walks through every path that actually works in 2026 — paid subscriptions, API access, and CLI tools — without a VPN and without a foreign credit card.
It assumes you are a developer, not someone looking for an unofficial workaround. Everything below is on the public, published path.
Why It's Hard From China
Two separate problems compound each other.
The network. chat.openai.com and api.openai.com are not reliably reachable from China Telecom, Unicom, or Mobile. Sometimes a request goes through; sometimes it times out at TLS handshake; sometimes the response stalls midway. This is not a fixed blocklist — it varies by region, time of day, and ISP routing. The only consistent solution is to route traffic through a stable endpoint that is reachable.
The billing. OpenAI requires a payment method that passes their fraud checks — typically a credit card issued in a country they support. Mainland China-issued Visa/Mastercard cards routinely fail. Domestic AMEX cards have intermittent success. This is not about paying more; many cards are rejected at the verification step before any charge happens. If the console itself refuses your region (unsupported_country_region_territory), see the region-block guide.
These two walls used to be solved by VPN + buying a US virtual card from a third party. That works, but it is fragile, often violates the third party's terms, and adds latency on every request. The 2026 path is simpler.
ChatGPT Plus: Three Paths That Work
If you only need the chat product (not the API), here are the three honest options:
| Path | Network reliability | Cost | Caveat |
|---|---|---|---|
| Personal VPN + foreign card | Variable | $20/mo + VPN cost | Card may still be rejected |
| Mainland-issued AMEX (US-style) | Good | $20/mo | Some banks decline; worth trying |
| Skip Plus, use API tools | Excellent (with Router One) | Pay-per-use, often cheaper | Different UX from chat.openai.com |
The third option is where most serious developers end up: pay for the API, run a local UI, use the same GPT-5.5 / Codex models — and skip the subscription entirely. A self-hosted chat UI such as LibreChat puts a ChatGPT-style interface on top of that key — its setup guide shows the custom-endpoint config. We cover that next.
The GPT API From China — The Path That Just Works
The cleanest 2026 path for the API is to point your code at an OpenAI-compatible endpoint that is directly reachable from China. Router One does exactly this: one set of credentials, the major model families the catalog currently lists — GPT (GPT-5.5, the GPT-5.6 models, GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol), Claude (Fable, Opus, Sonnet and Haiku ids), Gemini, Grok and DeepSeek — and RMB billing through Alipay in the hosted checkout. On subscription plans (as of the 2026-10-06 plan response), GPT-6.1 Sol, GPT-6 Sol, GPT-5.6 Sol and GPT-5.5 are in the Premium models tier of all three plans and GPT-6 Astra only in the Flagship models tier of Max and Ultra — the GPT-6.1 Sol API guide and the GPT-6 Sol API guide cover calling GPT-6.1 Sol and GPT-6 Sol.
The API surface is OpenAI-compatible, so any existing OpenAI SDK keeps working with two changes:
export OPENAI_BASE_URL=https://api.router.one/v1
export OPENAI_API_KEY=sk-your-router-one-key
That's it. from openai import OpenAI continues to work (OpenAI SDK). The Vercel AI SDK continues to work. LangChain continues to work. You get GPT-5.5 over a connection that does not require a VPN, with a balance you topped up using a method you already have. For full setup details see our Alipay top-up guide.
Router One measured 110-130 ms p50 across Beijing, Shanghai and Shenzhen in the published China latency benchmark; direct api.openai.com from the same networks, when it succeeds, is typically several hundred milliseconds with a long tail. The gap is large enough that even users with working VPNs sometimes prefer the routed path.
Codex CLI Without a Subscription
OpenAI ships Codex CLI as part of the broader Codex platform. The default flow expects you to sign in with a ChatGPT plan (Plus, Pro or a business plan) or use a billed OpenAI API account. Both are awkward from China for the reasons above.
Through Router One, Codex CLI works in pay-per-token mode — one command installs it, registers the provider in ~/.codex/config.toml, and verifies your key:
curl -fsSL https://router.one/install/codex.sh | bash -s -- sk-your-router-one-key
codex
You pay only for what the agent consumes — pay per token from your wallet, or use a live subscription plan for predictable monthly cost. No foreign card, no VPN. Based on the China latency benchmark last updated 2026-05-15, Router One measured 110-130ms p50 across Beijing, Shanghai, and Shenzhen; individual networks may vary. The setup, edge cases, and pricing breakdown are covered in the Codex CLI alternative guide, the Codex CLI without ChatGPT Plus guide and the China latency benchmark.
What About ChatGPT Plus Specifically?
Some developers really do want chat.openai.com itself — for image generation, voice mode, custom GPTs, and a few other consumer features that the API does not expose. There is no clean 2026 path to subscribe to ChatGPT Plus from a Chinese-issued card without intermediaries.
What does work, in order of preference:
- Mainland-issued multi-currency AMEX. Several Chinese banks issue AMEX cards (CMB, ICBC, Bank of China) with USD support. Acceptance rate at OpenAI's billing screen is variable but non-zero. Worth a try if you already have one.
- Family member abroad. A relative in the US, Singapore, Canada, or the UK who can subscribe and share access via OpenAI Team is the most stable solution. Team accounts are explicitly multi-seat and within ToS.
- Reputable virtual card services. Several services issue prepaid Visa/Mastercard for OpenAI billing. Acceptance fluctuates with OpenAI's fraud rules. Reliability is "OK most months, broken some months."
For developer use cases — code completion, agent loops, custom tools — the API path through Router One is materially better than any of the above. The reason to chase ChatGPT Plus is when you specifically need the consumer product surface.
Cost Comparison: Subscription vs API
A common worry is "API will be more expensive than $20/month." It depends entirely on how you use it.
| Profile | ChatGPT Plus ($20/mo) | GPT-5.5 API via Router One |
|---|---|---|
| Light chat use, ~50 messages/day | Better value at $20 flat | ~$5-12/month |
| Heavy chat use, 200+ messages/day | $20 (rate-limited) | ~$25-60/month |
| Daily code generation via Codex | Plus does not cover Codex CLI heavily | ~$30-100/month |
| Building agents that loop | Plus does not cover this | Pay per token from wallet or use a live subscription plan |
The non-obvious truth: for most developers, API usage is cheaper than the Plus subscription, because most developer chat sessions do not burn $20 worth of tokens in a month. The exception is heavy consumer use — image generation, long voice conversations — which the API does not cover the same way.
Setup in Five Minutes
If you want to try this end-to-end:
- Sign up at router.one with email or GitHub.
- Top up your balance with Alipay (no foreign card needed).
- Create an API key on the dashboard.
- Set the two environment variables above.
- Run any OpenAI SDK example. It works.
Switching back to direct api.openai.com later is a one-line change — Router One is OpenAI-compatible by design, so nothing in your code is locked in.
Comparing Router One to OpenRouter for China Use
OpenRouter is the closest international equivalent. The differences from a China-developer perspective:
- Network. Per its docs (checked 2026-10-01), OpenRouter offers a global endpoint, plus EU and US in-region endpoints on its Business and Enterprise plans, and documents no mainland-China endpoint. Router One is explicitly built for mainland reachability (China latency benchmark).
- Payment. Per its FAQ (checked 2026-10-01), OpenRouter accepts major credit cards, Alipay and USDC. Router One takes a card or Alipay through one hosted checkout, plus USDT and USDC on six chains (Tron, BSC, Ethereum, Polygon, Base, Arbitrum).
- Models. Both offer the major LLMs. The Router One catalog lists the GPT, Claude, Gemini, Grok and DeepSeek families — check router.one/models for the live lineup.
A full comparison is in our Router One vs OpenRouter for China post.
FAQ
Does using Router One violate OpenAI's terms? You call Router One's API under Router One's terms of service. Model providers' usage policies still apply to the content you send, and using the service is not a license to bypass the law, sanctions or upstream model terms — you need an eligible jurisdiction and a compliant workload. For contractual questions, contact support.
Can I use my existing OpenAI API key with Router One? No. Router One issues its own keys. The point is to skip OpenAI's billing entirely; using your own key would defeat that. If you already have an OpenAI account that works, you can keep using it — Router One is for cases where direct access is blocked or impractical.
What about data privacy? Router One does not retain prompt or model-response bodies from direct API calls. It records request metadata for billing, usage tracking, and troubleshooting. Playground saves conversation history so users can revisit and continue their chats. Read data retention and the privacy policy, and confirm applicable upstream data-use terms with support.
Does this work for ChatGPT Team / Enterprise? Team and Enterprise are subscription products from OpenAI; Router One is API access. They are different shapes of relationship. If your organization needs Team for collaboration features, that's a separate purchase.
What if a model I need is not on Router One?
Most users will find what they need — GPT-5.5, the GPT-5.6 models, GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol, the Claude family (Fable, Opus, Sonnet and Haiku ids), Gemini 2.5 Pro and the Gemini Flash line, Grok, and DeepSeek (deepseek-v4.1-flash and deepseek-v4-flash) are on the platform; the live lineup is on router.one/models. If something is missing, contact us and we'll evaluate adding it.
Will OpenAI block this kind of access? Router One cannot speak for OpenAI's future policies. What it does control: your request is served by the model you name, a failing route (429, 5xx or timeout) is retried on another route for the same model, and active incidents are posted on /status.
Conclusion
The 2026 reality for Chinese developers is that the API path is more useful than chasing ChatGPT Plus. Sign up at router.one, top up with a card or Alipay, point your code at the OpenAI-compatible endpoint, and you have GPT-5.5, Codex CLI and the other GPT models in the catalog with no VPN and no foreign card. Based on the China latency benchmark last updated 2026-05-15, Router One measured 110-130ms p50 across Beijing, Shanghai, and Shenzhen; individual networks may vary. For more on the underlying routing architecture, see AI model routing explained; for Cursor-specific China issues, see the Cursor Pro China guide.