Prepay credit, choose a reviewed Bedrock model, and set a workload budget. Managed requests stop when eligible credit cannot cover the required reserve.
AI Gateway HQ maintains a reviewed catalog of U.S. conversational models. Model availability is checked before it appears in the customer catalog. Region, API support, output limit, and the current price source must pass release checks. A later check can remove a model that no longer passes.
02
Reserve credit before inference
Before inference, the gateway reserves settled credit for the allowed request. It then replaces the reserve with the measured token charge. No eligible settled credit means no provider call.
03
Do not add shared conversation state
Each request stays bound to its organization, workload, route, budget, and receipt. The adapter sends that request to Bedrock independently. It does not add shared memory between customers.
Current shared controls
What every supported connection includes
No customer AWS credential on the managed path Stable model aliases and reviewed route targets Organization and key budgets, rate limits, and concurrency limits Observe, Shadow, and Enforce modes with metadata-only evidence
Test the exact model, region, request format, streaming behavior, and displayed price before production approval.
Applications keep a stable gateway URL and model alias for Amazon Bedrock. Administrators can change eligible managed models behind the alias.
Implemented boundary
Supported protocols
Add settled credit and keep a reusable card or Stripe Link authorization on file. Choose a reviewed model, attach it to a workload route, and set its budget. No customer AWS credential is required.
POST /v1/responses through Bedrock Converse POST /v1/chat/completions (non-streaming) POST /v1/messages (non-streaming)
Compatibility follows the provider's current published interface. Review the provider documentation before approving Amazon Bedrock for production.