
Why Claude Code and Codex Fail at Hong Kong Airport
Claude Code and Codex both failed to sign in at Hong Kong International Airport. It wasn't the Great Firewall — here's what was actually blocking them, with sources
Sitting at a gate in Hong Kong International Airport, I opened Claude Code like I always do. My account showed up in the bottom-left corner, but sign-in kept failing. Codex did the same thing. My first guess was the Great Firewall. That guess was wrong.
It's not the Great Firewall
The Great Firewall only applies to mainland China, not Hong Kong. The airport's free Wi-Fi (HKAirport Free WiFi) worked fine for everything else — news sites, social media, all of it loaded normally.
So the block wasn't government censorship from Beijing or Hong Kong. It came from Anthropic and OpenAI themselves, as a deliberate access policy.
Anthropic deliberately excludes Hong Kong and mainland China
Anthropic's official supported-regions list doesn't include Hong Kong or mainland China. What makes this particularly unforgiving is that Claude checks IP location in real time, on every browser/client session — not just at login. ChatGPT, by contrast, only checks IP at login time, which is why it sometimes keeps working mid-session even from a restricted region.
That matches exactly what I saw: the desktop app still shows a cached account in the corner, but the actual authentication request gets rejected by IP geolocation.
Anthropic's stated reasons come down to two things:
- Legal uncertainty around data-sharing obligations under Hong Kong's National Security Law
- An expansive reading of US export control regulations (EAR) covering the Asia-Pacific region
In September 2025, the restriction went global
Digging further, I found that Anthropic expanded this policy in September 2025 to block China-controlled entities worldwide — not just users physically in Hong Kong or mainland China ( SCMP ). Any company more than 50% owned, directly or indirectly, by entities in unsupported regions is affected, regardless of where it's incorporated. One industry lawyer called it the first time a major US AI company had issued this kind of formal, public prohibition.
While I was researching this, a news story broke that made it feel oddly current: OKX, the crypto exchange, banned Claude access for its Hong Kong-based staff ( Crypto Briefing ). It started with a corporate account suspension scare in early August. The account was eventually restored, but OKX pulled Claude access for employees in Hong Kong and mainland China afterward, and told staff internally that using a VPN to route around the block would not be supported by the company either. Goldman Sachs reportedly follows the same restriction.
Codex has the same problem, same root cause
OpenAI isn't any better here. It has deliberately blocked ChatGPT and API access from Hong Kong and mainland China since July 9, 2024, citing US export control compliance ( OpenAI Help Center ). Since ChatGPT only checks IP at login, some sessions started elsewhere may keep working temporarily, but Codex hits the API on every call, so it gets blocked reliably from a Hong Kong IP.
Chinese-origin models (Qwen, Kimi) don't have this problem
I got curious whether this applies to Chinese models like Qwen (Alibaba) or Kimi (Moonshot AI). It doesn't — if anything, these are built assuming users in Hong Kong and mainland China, so regional restrictions are effectively absent.
US export controls target the transfer of technology from US companies to China — they don't touch Chinese companies' own services. Qwen (Apache 2.0) and DeepSeek are accessible with no regional gating, and Kimi's consumer-facing product is reportedly operated through a Singapore entity even though Moonshot AI develops it. Same category of tool, opposite regulatory treatment, depending on which side of the export-control line the vendor sits on.
The mismatched adult-site blocking is a separate story
As a side note, I noticed some sites loaded fine over the airport Wi-Fi while others didn't. That's unrelated to the AI restrictions above. It's almost certainly the airport's own DNS-based category filtering, which is standard for public Wi-Fi operators blocking adult/gambling categories ( DNSFilter ). A 2021 study that tested 1,249 domains across US airport Wi-Fi networks found this kind of filtering is often inconsistent, with real gaps in coverage. The uneven blocking I saw is more likely a gap in the filter's domain database than any kind of political censorship.
Does Amazon Bedrock get around this?
Since Bedrock access to Claude is a contract with AWS rather than a direct contract with Anthropic, I wanted to know whether the geo-block logic changes — and whether picking Tokyo, Seoul, Singapore, or Taipei as the region matters.
> Access to Anthropic models is not allowed from unsupported countries, regions, or territories
> your AWS account and billing address must also be in a supported country, region, or territory
So Bedrock does layer an extra gate on top, specific to Anthropic models. The rest of AWS — console sign-in, other model providers — works fine with a Hong Kong billing address; Anthropic models alone carry this additional restriction.
The gating factor appears to be primarily the AWS account's billing/tax address, not real-time IP checking on every request the way claude.ai does — though I couldn't find a solid primary source confirming or ruling out live IP checks on Bedrock's side. One data point suggesting the restriction logic isn't fully documented: a GitHub report describes a user in Pakistan — a country that is on Anthropic's supported list — getting blocked through Bedrock anyway ( pollinations/pollinations#10613 ), with the note that "Bedrock enforces additional geographic restrictions beyond what Anthropic's native API requires."
There's also a practical gotcha worth knowing: some Claude Code CLI setups configured to use Bedrock as the backend still make background calls (telemetry, usage tracking, oauth) directly to
api.anthropic.com ( openclaw/openclaw#30672 ). In that case, model inference itself might succeed through Bedrock while the tool still errors out because of that side-channel call getting geo-blocked.As for region choice — Tokyo, Seoul, Singapore, Taipei — it turns out not to be the deciding factor. These just determine where the model is hosted; they don't change the geo-block itself, which is tied to account registration rather than region. So:
- An AWS account registered with a Hong Kong billing address will likely get blocked the same way regardless of which of these regions you pick
- An AWS account registered in a supported country (e.g. Japan, Singapore) will likely work across any of these regions
Taipei (ap-east-2) is a newer region launched in 2025 and appears to lag slightly behind Tokyo/Singapore in model availability.
Wrap-up
I started this investigation assuming "Hong Kong" automatically meant "Great Firewall." It turned out the actual blocker was US export-control compliance decisions made by American companies, layered independently of anything the Hong Kong or Beijing government does. And mid-investigation, the OKX story landed — proof this isn't just an airport-Wi-Fi inconvenience but something companies are actively navigating right now.
I haven't personally tested Bedrock's Anthropic-model access from a Hong Kong IP, so that part stays a documented-but-unverified gap. If anyone has run that test, I'd genuinely like to know the result.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article