The inference_geo parameter on the Claude API controls where your inference is processed. Set it to "us" and processing stays in US data centers; leave it at the default "global" and Anthropic can serve the request from any of its data centers. It exists for one reason: data residency. It is not a capacity knob — and Anthropic's rate-limit documentation says so directly: rate limits apply per model and are shared across inference_geo values. Requests tagged "us" and requests tagged "global" draw down the same token bucket.
Why the confusion is understandable
On the cloud platforms, geography and capacity genuinely are entangled. On Vertex AI, the global endpoint and the us/eu multi-region endpoints have independent quota buckets. On Foundry, Global Standard and Data Zone Standard deployments draw from different pools. Someone who learned multi-region capacity math on those platforms might reasonably assume the Claude API's inference_geo works the same way. It does not: on the first-party API (and on Claude Platform on AWS, which supports the same parameter), residency routing is a flag on the request, not a separate deployment with its own allocation. Your organization's RPM, ITPM, and OTPM limits for a model are what they are, however you split the geography.
What inference_geo actually changes
Three things — none of them throughput:
Where inference runs. "us" pins processing to US data centers; "global" (the default) allows any Anthropic-operated data center.
What you pay. For Claude Opus 4.6, Sonnet 4.6, and later models, inference_geo: "us" applies a 1.1x multiplier on all token pricing categories — input, output, cache writes, and cache reads. Global routing uses standard pricing. So the "split the pool" idea is worse than neutral: the US-tagged half of your traffic costs 10% more for exactly the same rate-limit consumption.
Which models accept the request. Models released before Opus 4.6 and Sonnet 4.6 return a 400 error if the parameter is included at all.
On Claude Platform on AWS there is one more subtlety worth naming: the AWS region of your workspace controls the gateway endpoint and where AWS-side resources like IAM and CloudTrail are scoped — it does not pin where model inference runs. Only inference_geo (per request, or a workspace-level default_inference_geo) does that. Workspace region vs. inference geography unpacks this distinction.
from anthropic import AnthropicAWS
client = AnthropicAWS() # needs AWS_REGION + ANTHROPIC_AWS_WORKSPACE_ID
resp = client.messages.create(
model="claude-opus-4-8",
max_tokens=512,
inference_geo="us", # residency + 1.1x pricing; same rate-limit pool
messages=[{"role": "user", "content": "Summarize this contract..."}],
)Verifying where traffic actually ran
If you adopt inference_geo for compliance reasons, verify it with data rather than trust. The Usage Admin API can group and filter usage by inference_geo (values global, us, or not_available for pre-February-2026 models), which gives you an auditable record that residency-tagged traffic was routed as intended. Details in the Usage API dimensions guide.
So how do you actually get more capacity?
Since geography splitting is a dead end on the first-party API, the real levers are the usual ones: moving up usage tiers (or negotiating a Custom tier), prompt caching — cache reads do not count toward ITPM on most models, which multiplies effective throughput — offloading asynchronous work to the Batch API, which has its own separate limits, and spreading across genuinely independent pools such as different model families or different platforms. Note that inference_geo itself is a first-party and Claude Platform on AWS feature; Bedrock and Vertex AI do not support the parameter, and on Foundry the equivalent is choosing a US Data Zone Standard deployment, which Anthropic documents as equivalent to inference_geo: "us" including the same 1.1x multiplier.
Where to go next
Quotas and Rate Limits covers how the shared pools are metered, and the pooled-vs-isolated guide shows where multi-region accounting genuinely does add capacity on the cloud platforms.