
I Tested Bedrock From Hong Kong Across 4 Regions Description
I went back through Hong Kong Airport and actually tested Claude, GPT, and other models on Bedrock across four regions. Here's what got blocked and what didn't.
In my previous post , I wrote about Claude Code and Codex failing to sign in at Hong Kong International Airport, and ended with an open question about Amazon Bedrock, based only on reading AWS's documentation. On August 23, I passed through the same airport again on a connecting flight home, so I actually ran the test this time: four Bedrock regions, Playground, several model families.
What I did
Connected to the airport's Wi-Fi, signed into the AWS Bedrock console, and sent a simple test prompt to models in Tokyo, N. Virginia, Singapore, and Taipei. The AWS account's billing address is registered to my home address in Tokyo.
Results
Target | Tokyo | N. Virginia | Singapore | Taipei |
|---|---|---|---|---|
Anthropic Claude (Haiku 4.5 / Sonnet 5 / Opus 4.5) | Blocked | (untested — different error) | Blocked | Blocked |
OpenAI GPT-5.6 series (proprietary) | Blocked | — | Blocked | — |
gpt-oss-120b (OpenAI open-weight) | Worked | — | — | — |
Kimi K2.5 (Moonshot AI) | Worked | — | — | — |
Nova Pro (Amazon's own model) | — | — | — | Worked |
platform.claude.com (Anthropic's own domain) | Blank page | Blank page | — | — |
Claude Sonnet 5 in N. Virginia and Singapore returned "The provided model identifier is invalid" — a different kind of error, just meaning that particular model ID isn't valid in that region. The actual geo-block error is a
ValidationException reading "Access to Anthropic models is not allowed from unsupported countries, regions, or territories" — the exact wording documented on AWS re:Post .OpenAI models return the same kind of error, worded "Access to OpenAI models is not allowed from unsupported countries, regions, or territories," pointing to OpenAI's own supported-countries page. So Bedrock passes through each provider's own regional restriction rather than replacing it with something uniform.
Tokyo

N. Virginia

Billing address in Tokyo didn't matter
This is the clearest finding from the trip. My previous post left open whether Bedrock's geo-block is based on the AWS account's billing address or real-time IP detection. This account's billing address is registered in Tokyo — a fully supported region. It still got blocked on every Anthropic model call, every time, the instant I connected from Hong Kong Wi-Fi.
So Bedrock does check connection origin, not just account registration. A supported billing address alone isn't enough to guarantee access.

Switching regions didn't help either
Tokyo, Singapore, and Taipei all produced the identical block on Anthropic models. My previous post guessed that region choice wasn't the deciding factor and that account registration was what mattered — that turned out to be only half right. Account registration alone doesn't save you either; picking a different region is not a workaround.
Only proprietary models get blocked
The clean split by model type was the other big finding.
- Anthropic Claude (Haiku 4.5, Sonnet 5, Opus 4.5) — blocked across the board
- OpenAI GPT-5.6 series (proprietary hosted models) — blocked across the board
- OpenAI's gpt-oss-120b (open-weight) — answered normally from Tokyo
- Kimi K2.5 (Moonshot AI) — answered normally from Tokyo, even introducing itself as "Kimi, functioning normally"
- Nova Pro (Amazon's own model) — answered normally from Taipei
Same Bedrock console, wildly different treatment depending on how the model is delivered. The regional restrictions Anthropic and OpenAI apply to their own closed models don't extend to open-weight models, Amazon's in-house models, or Chinese-origin models. My previous post's prediction — that Chinese models sit outside US export control scope and would work fine — held up exactly as expected once I actually tested Kimi.



platform.claude.com stayed blank in both Tokyo and N. Virginia
One more thing worth flagging: the AWS console's "Claude on Bedrock" panel links out to
platform.claude.com/partner-login . That page rendered completely blank over Hong Kong Wi-Fi, regardless of whether the Bedrock console was set to Tokyo or N. Virginia.Since switching Bedrock's region didn't change this page's behavior at all, it's most likely a separate block happening at the
platform.claude.com domain level — Anthropic's own infrastructure, independent of Bedrock's regional model access logic. This lines up with something I mentioned in the last post: some Claude Code CLI setups configured to use Bedrock still make background calls straight to api.anthropic.com for telemetry and oauth ( openclaw/openclaw#30672 ). Routing model inference through Bedrock doesn't fully insulate you from Anthropic's own domains.
The Bedrock console itself also threw an unrelated "Something went wrong. Try reloading the page" error partway through. That's a different message from the geo-block wording, so I can't tell whether it was region-related or just a transient glitch.

Wrap-up
Three things came out of this trip:
- Bedrock passes through each provider's own geo-restriction on their proprietary models, and it checks connection origin, not just account registration. A Tokyo billing address didn't help.
- Switching between Tokyo, Singapore, and Taipei made zero difference. Region selection is not a workaround.
- Only closed, proprietary models get blocked. Open-weight models (gpt-oss), Amazon's own models (Nova), and Chinese-origin models (Kimi) all worked fine.
I closed the last post saying I'd like to know if anyone had actually run this test. Turns out it was going to be me, on the same layover, a few days later.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article