Grok 4.6 buying guide: The Bedrock calling region is not necessarily the inference region

Use three fictional tickets to choose between Oregon Mantle and keeping Converse on Runtime. Compare data boundaries and validation costs, not just token rates.

Source: aws.amazon.com
November 17, 2026
3 min

A team calling Bedrock from Northern Virginia cannot assume inference happens there. When selecting Grok 4.6, separate the regions allowed to process the data from the response contract. Do you need the service to constrain the format, or can the application validate it? If both are firm requirements, check that the same route meets both before looking at prices.

What each endpoint lets you keep

AWS’s launch article describes two Bedrock endpoints for the same xAI model. The Grok 4.6 model card lists in-Region inference on commercial AWS bedrock-mantle in Oregon (us-west-2), with JSON Schema structured outputs. If your AWS SDK pipeline already uses Converse, bedrock-runtime keeps that interface. For Grok 4.6 it is cross-Region only, however, and lacks the same service-enforced structured output. Even a US Geo call from Northern Virginia can be routed to Ohio or Oregon as well. This comparison covers commercial AWS, not GovCloud.

Decide with three paper test cases

Use three fictional repair tickets. Only ticket_id and category are allowed, and the latter must be power, display, or other. Draft one normal response, one with an extra field, and one with category battery. Before calling any model, write down the permitted destinations and verify that your parser accepts the first response and rejects the other two for human review. Sketch both proposed configurations: Mantle with Oregon and server-side Schema; Runtime with Converse, every possible destination for the selected inference profile, and application validation. Enforced format does not prove classification accuracy; still check the ticket ID and the reason for its category.

The AWS pricing page links to the model card for Grok 4.6 rates. Assume the Standard tier, 10 million uncached input tokens and 1 million output tokens, with no retries or other charges. In-Region and Geo input/output rates are USD 2.20/6.60 per million, totaling USD 28.60. Global is USD 2.00/6.00, totaling USD 26.00. This is a paper estimate, not a bill or a measured test. Global’s lower rate only matters if policy permits its destinations and the team can afford validation, rejection handling, and review.

If inference must stay in Northern Virginia and Schema must be enforced by the service, or if you require both Converse and server-side Schema, neither endpoint is a direct substitute: keep the existing implementation or verify another model against both constraints. If Oregon is allowed and the output contract matters, investigate Mantle first. If Converse must stay, assess Runtime only when its cross-Region destinations and your fallback pass review.

Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.

Original source:aws.amazon.com