[DOCS] Gateway "Different regions" lever contradicts the models section for non-US Bedrock geos; ANTHROPIC_BEDROCK_REGION_PREFIX undocumented

Documentation Type

Incorrect/outdated documentation

Documentation Location

https://code.claude.com/docs/en/claude-apps-gateway-config

Section/Topic

upstreams → Amazon Bedrock (the "Different regions" lever), and the
undocumented ANTHROPIC_BEDROCK_REGION_PREFIX env var on the Bedrock client
page.

Current Documentation

Two places, one wrong and one missing.

1. The "Different regions" lever in the multi-upstream table says:

Different regions — One Amazon Bedrock upstream per region, each with its own region:. With auto_include_builtin_models: true the cross-region inference profiles route automatically; for region-pinned deployments use a models: block.

2. ANTHROPIC_BEDROCK_REGION_PREFIX appears in the changelog but in no docs
page. It is not on the Bedrock client page, the gateway config reference, or
the settings/env-var reference.

What's Wrong or Missing?

On the "Different regions" lever: "the cross-region inference profiles route
automatically" is only true inside the US geo. The built-in catalog's Bedrock
ids are all us.anthropic.*, and a geographic us.* profile does not route
across the EU or APAC geos. So an eu-west-1 or ap-northeast-1 upstream with
auto_include_builtin_models: true sends us.anthropic.… to
bedrock-runtime.<region>.amazonaws.com and the request fails — the opposite of
routing automatically. The sentence also frames a models: block as needed only
for "region-pinned deployments," when it is needed for any non-US geo.

The models
section already states this correctly — *"It is required for non-US Amazon
Bedrock regions"* — so the two sections contradict each other, and the one an
operator hits first while sizing a multi-region deployment is the wrong one.

On ANTHROPIC_BEDROCK_REGION_PREFIX: none of its behavior is discoverable
without reading the binary. Specifically undocumented:

• Its accepted values: us, eu, apac, jp, au, global. It is validated

against that set, so a typo or an unlisted value is silently dropped.

• It takes precedence over the prefix otherwise derived from AWS_REGION

(us-* → us, eu-* → eu, ap-* → apac, anything else → global).

• us-gov-* regions ignore it entirely and are forced to the us-gov prefix.
• It is checked against ListInferenceProfiles at startup, and models with no

matching profile in the account fall back to the AWS_REGION-derived prefix
with a warning. If that list call fails, the prefix is applied unchecked and a
different warning says requests may 400.

Suggested Improvement

1. Reword the "Different regions" lever to match the models section, e.g.:

*"One Amazon Bedrock upstream per region, each with its own region:. The
built-in catalog uses US geographic inference profiles (us.anthropic.*), so
non-US geos and region-pinned deployments both require a models: block with
the right per-upstream IDs."*

1. Document ANTHROPIC_BEDROCK_REGION_PREFIX on the Bedrock page with the

accepted values, the precedence rule, the us-gov exception, and the
availability check.

Impact

Both cost real debugging time on a multi-region Bedrock rollout. The lever table
reads as "regions are handled for you," so the failure surfaces as an opaque
Bedrock 400 on the first request from a non-US upstream. And an operator who
finds ANTHROPIC_BEDROCK_REGION_PREFIX in the changelog has no way to learn its
values or its scope short of running strings on the release binary — which is
how the list above was obtained.

Additional Context

Verified against the 2.1.224 native binary (darwin-arm64, GIT_SHA
8a2a469b68f918917492973f3b16bd1682b9f82c).

Filed alongside #84875, which covers the underlying gap: there is no supported
way to point a gateway at global.* inference profiles.