Setting inference_geo: "us" keeps model inference inside US-based infrastructure — and costs 1.1x standard pricing across every token category. Here's exactly what it restricts, what it doesn't, and how it interacts with caching and Priority Tier.
inference_geo | Pricing | What it means |
|---|---|---|
"global" (default) | 1x — standard rates | Inference may run in any available geography for best performance/availability. |
"us" | 1.1x — all token categories | Inference runs only in US-based infrastructure. |
The 1.1x applies uniformly to input tokens, output tokens, cache writes, and cache reads — not just the input side. A 1-hour cache write at $10/MTok (Opus 4.8) becomes $11/MTok under inference_geo: "us"; a cache-hit read at $0.50/MTok becomes $0.55/MTok. There's no separate "US output rate" table — it's a flat multiply on whatever the base rate already is.
Inference geo (the priced one) controls where the model actually runs, per request via the inference_geo parameter or a workspace default. Workspace geo is a separate, unpriced setting fixed at workspace creation — it controls where data is stored at rest and where endpoint processing (image transcoding, code execution) happens. Right now "us" is the only workspace geo option that exists, so this second setting isn't really a choice yet — the pricing story is entirely about inference_geo.
| Model generation | inference_geo support |
|---|---|
| Opus 4.6, Sonnet 4.6, and later | Supported — 1.1x applies to "us" |
| Opus 4.5, Sonnet 4.5, Haiku 4.5, and earlier | Not supported — including the parameter returns a 400 error |
Also not applicable: Amazon Bedrock and Google Cloud (inference region is set by the endpoint URL/inference profile instead), the OpenAI SDK compatibility endpoint, and Claude Managed Agents (which respects Workspace geo but doesn't accept the per-request parameter at all).
| Modifier | Interaction with inference_geo: "us" |
|---|---|
| Prompt caching multipliers | Stack — 1.1x multiplies the already-adjusted cache write/read rate |
| Fast mode premium | Stacks on top of fast mode rates |
| Batch API | Supported — each request in a batch can set its own inference_geo |
| Priority Tier commitment | Burns down faster — each US-routed token draws 1.1 tokens from committed TPM, same mechanism as caching multipliers |
That Priority Tier interaction is easy to miss: it's not an extra dollar cost on top of the 1.1x token price, it's a second effect where your committed-capacity pool depletes 10% faster whenever requests are pinned to "us".
On Claude Platform on AWS, the same 1.1x applies to inference_geo: "us". On Claude in Microsoft Foundry, the equivalent is the US Data Zone Standard deployment type on Azure, priced the same 1.1x. Bedrock and Google Cloud are different: region is picked via endpoint URL, not this parameter, and each has its own independent regional pricing — check Bedrock pricing or Google Cloud pricing directly rather than assuming the 1.1x carries over.
Anthropic migrated legacy "US-only" organization opt-outs automatically to allowed_inference_geos: ["us"] with default_inference_geo: "us" — same enforced behavior, no code changes required, and no need to add inference_geo to every call manually. If your residency requirements have since relaxed, switching back to global (and standard pricing) means updating the workspace's allowed/default geo settings in Console or the Admin API, not just dropping the parameter from requests.
The 10% premium is the cost of a real constraint, not a performance feature — global routing exists specifically for "best performance and availability," so opting into "us" trades some of that away deliberately for a compliance/contractual requirement. If nothing forces US-only inference on you, there's no reason to pay it. If something does (a customer contract, a regulatory requirement), 1.1x is a fixed, predictable line item — budget it into whichever rate you're already using, and remember it also compounds with fast mode and cache pricing rather than replacing them.
Use the full calculator for standard-rate estimates, or see rate limits for how Priority Tier commitments and TPM burndown work in more depth.