The double block — why a VPN alone often isn't enough
ChatGPT enforces two walls for China users. Wall 1 is the GFW: chatgpt.com, chat.openai.com, and platform.openai.com are DNS-poisoned + IP-blocked, direct access times out. Wall 2 is OpenAI's own region denial — even with a VPN exit, OpenAI re-verifies the IP against its supported-country allowlist (~180 countries); China, Russia, Iran, NK, Cuba, Syria IPs get 'Access Denied. You are accessing from a country we do not currently support'. That means even an HK, Ukrainian, or Russian exit — geographically close but on the denylist — still fails.
OpenAI's IP identification mechanism
OpenAI's anti-VPN is two-layer: Cloudflare + in-house. Layer 1 — Cloudflare Bot Management identifies datacenter IPs at the edge (AWS / GCP / Azure / OVH) and serves low-quality IPs either a human-check or a refusal. Layer 2 — OpenAI maintains its own country denylist; IPs in those countries get refused outright. Layer 3 — account-level behavioral detection; logging in from multiple countries in 24h triggers a phone-number re-verification. Practical fix: pick a residential IP in a supported country (US / JP / SG) rather than a pure datacenter VPN exit — hit rate rises sharply.
AF3 node recommendations — stable US / JP / SG exits
AF3 rates ChatGPT at 1-star strictness — as long as the IP isn't on the denylist and isn't an obvious datacenter, you're usually fine. Preferred: US-West (California / Oregon — closest to OpenAI's us-west deployment, 120ms usable), Tokyo (50ms, consistent quality), Singapore (best for SEA). Avoid: Hong Kong (on OpenAI's denylist), Russia, Turkey, and some LatAm small countries (occasional refusals). Transport layer: use TLS 1.3 + ECH or Reality protocols to avoid SNI leaks that trigger GFW detection. Extra hardening: clear Cloudflare cookies, disable WebRTC in the browser.
Plus signup and payment — US Apple ID is the simplest path
ChatGPT Plus ($20/mo) and Pro ($200/mo) require: (1) an overseas phone number (SMS-Activate / 5Sim — but OpenAI is rapidly blacklisting bulk SMS ranges); (2) overseas payment (credit card or Apple ID). The cleanest path: a US Apple ID topped up with US gift cards, subscribing through iOS — App Store payments don't validate billing address, only Apple ID country. Caveats: one phone per account, re-rolling is expensive; switching payment country too often triggers account risk; Chinese email domains (@qq.com / @163.com) work but Gmail / Outlook raise fewer captchas.
Ban avoidance and account security
OpenAI bans on a handful of triggers: (1) multi-country login in a short window (VPN flip-flopping); (2) heavy concurrent API calls past quota; (3) prohibited content (malware generation, CSAM, etc.); (4) multi-account sharing the same payment or phone number. Field practice: keep the same exit country and the same browser fingerprint for login; use your Plus monthly $20 API credit rather than re-billing; avoid using one US card for 3+ account renewals. Appeals are rarely successful — back up your chat history proactively. You can't casually swap your account's region; doing so drops chat history (except workspace-logged threads).
ChatGPT vs. Claude vs. Gemini — access-difficulty comparison
All three region-deny mainland China, but tolerances differ. ChatGPT: double block; needs a working VPN in a supported country. Claude: Anthropic's denylist is stricter — adds parts of Thailand, Vietnam, and certain Brazilian ASNs compared to OpenAI's, and is more sensitive to CloudFront datacenter exits. Gemini: easier once signed into Google, because Google leans on reCAPTCHA + account history rather than hard IP refusal — though the Gemini app is pulled from Chinese stores. Ranking by VPN tolerance: Gemini > ChatGPT > Claude. Domestic alternatives exist — Ernie (Baidu), Qwen (Alibaba), DeepSeek — but training-data and ecosystem gaps are significant.