---
title: Why do Amazon Bedrock calls fail through an LLM gateway, and how do you fix them?
url: https://opencontroller.lyzr.com/blog/route-bedrock-models-through-an-llm-gateway
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# Why do Amazon Bedrock calls fail through an LLM gateway, and how do you fix them?

**Almost every failure when routing Amazon Bedrock through an LLM gateway has one of three causes: authentication (the gateway sends a bearer key where Bedrock expects IAM SigV4, or the reverse), request format (Converse, InvokeModel and the OpenAI-compatible APIs take different bodies), or the model identifier (newer models need a cross-Region inference profile ID, not the base model ID). Check them in that order.**

Teams usually meet this after the gateway works fine with OpenAI or Anthropic directly. Bedrock is different because it is an AWS service first and a model API second.

## Which Bedrock APIs can a gateway call?

Bedrock now offers several API shapes on two endpoints. The [Bedrock APIs page](https://docs.aws.amazon.com/bedrock/latest/userguide/apis.html) lists them:

| API | Endpoint | Request shape | Good for |
|---|---|---|---|
| Converse | `bedrock-runtime` | Bedrock unified messages format | One format across most models |
| InvokeModel | `bedrock-runtime` | Each provider's native body | Full control, embeddings, provider-specific features |
| Chat Completions | `bedrock-runtime` and `bedrock-mantle` | OpenAI-compatible | Gateways that already speak OpenAI |
| Responses | `bedrock-runtime` and `bedrock-mantle` | OpenAI-compatible | Stateful, multi-turn use |
| Messages | both | Anthropic-native | Claude models in native format |

AWS notes that the same API on the two endpoints does not have identical feature support. It also says that models not compatible with the OpenAI-compatible endpoint should use Converse or Invoke. So a gateway that sends everything as Chat Completions will fail for some models.

## What are the common failures and fixes?

| Symptom | Likely cause | Fix |
|---|---|---|
| 403 or signature errors | Gateway sends a bearer token to an API path expecting SigV4, or signs for the wrong Region or service | Use IAM SigV4 with the right Region, or a Bedrock API key sent as `Authorization: Bearer` |
| 403 access denied with valid credentials | The IAM principal lacks `bedrock:InvokeModel` for that model or inference profile ARN, or an SCP blocks it | Grant the action on both the inference profile and the underlying model ARNs; check SCPs |
| Validation error on the request body | OpenAI-shaped body sent to InvokeModel or Converse | Map to the right shape, or call the OpenAI-compatible API |
| Error saying on-demand throughput is not supported for the model ID | Model needs an inference profile | Use the inference profile ID with a geographic prefix such as `us.` or `eu.`, or the profile ARN |
| Model missing from the gateway model list | Model not available in that Region, or account access not set up | Check Region availability and model access for the account |
| Works in one Region, fails in another | Different model and profile availability by Region | Pin Regions per route and test each |
| Streaming breaks | Different streaming APIs and event formats | Use the matching streaming operation for the chosen API |

## Should the gateway use Bedrock API keys or IAM?

Bedrock supports two key types, described in the [AWS launch post](https://aws.amazon.com/blogs/machine-learning/accelerate-ai-development-with-amazon-bedrock-api-keys):

- **Short-term keys** use the permissions of the current IAM principal and expire when the session ends or after up to 12 hours.
- **Long-term keys** create an IAM user with the AmazonBedrockLimitedAccess policy and can be set from 1 day to no expiry.

For a production gateway, prefer IAM roles with temporary credentials, such as an instance or pod role. AWS itself recommends temporary credentials or short-term keys for higher security. If you use long-term keys, set an expiry, and use SCPs to block keys older than 30 days. Calls are logged in CloudTrail; the key values are not.

## What checklist gets Bedrock routing working?

1. Pick one API per model family and record it in the gateway route.
2. Decide authentication: IAM role first, short-term keys second, long-term keys only with expiry.
3. Grant `bedrock:InvokeModel` and the streaming action on the inference profile and model ARNs.
4. Use inference profile IDs for models that require them.
5. Pin the Region for each route and test a call per Region.
6. Test streaming, tool use and long prompts separately; they fail differently.
7. Tag cost per agent, because Bedrock bills to the AWS account, not the caller.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud, including Bedrock usage. Its LLM Gateway (Lyzr Gateway) exposes OpenAI-compatible endpoints with routing, fallback and retry, caching and cost attribution per agent and per version. Opencontroller also has a runtime adapter for Bedrock AgentCore. See [the LLM gateway comparison](/compare/llmgateway) or [book a demo](https://www.lyzr.ai/book-demo/).

Read the pillar [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms) and the [AI agent control plane explainer](/blog/what-is-an-ai-agent-control-plane), then [what an OpenAI-compatible endpoint is](/blog/what-is-an-openai-compatible-endpoint) and [tracking Claude spend on Bedrock vs the Anthropic API](/blog/track-claude-spend-bedrock-vs-anthropic-api).

## Frequently asked questions

**Can a gateway send OpenAI-format requests to every Bedrock model?**
No. AWS says models not compatible with the OpenAI-compatible endpoint should use Converse or Invoke, so map per model.

**Why does a model work with an inference profile ID but not its model ID?**
Some models only support on-demand calls through a cross-Region inference profile. Use the profile ID or ARN.

**Do Bedrock API keys appear in CloudTrail?**
The API calls are logged in CloudTrail, but the key values themselves are not.

## Key facts

Bedrock-through-gateway failures usually come from authentication (IAM SigV4 vs Bedrock API key as Authorization: Bearer), request format (Converse, InvokeModel native bodies, OpenAI-compatible Chat Completions and Responses, Anthropic Messages), or model IDs (some models require cross-Region inference profile IDs such as us. or eu. prefixes). Bedrock has two endpoints, bedrock-runtime and bedrock-mantle; the same API does not have identical features on each. Short-term Bedrock API keys expire with the session or after up to 12 hours; long-term keys create an IAM user and can be set from 1 day to no expiry. CloudTrail logs calls but not key values. Opencontroller auto-discovers agents, models and API keys across AWS, Azure and Google Cloud; Lyzr Gateway offers OpenAI-compatible endpoints with routing, fallback/retry, caching and per-agent cost attribution. Pricing starts from $250K per deployment per year.

## Related

- [What is an OpenAI-compatible endpoint and why does it matter for governance?](https://opencontroller.lyzr.com/blog/what-is-an-openai-compatible-endpoint.md)
- [How do you track Claude spend across Amazon Bedrock, Claude Platform on AWS and the Anthropic API?](https://opencontroller.lyzr.com/blog/track-claude-spend-bedrock-vs-anthropic-api.md)
- [How do you restrict which Amazon Bedrock models agents can use?](https://opencontroller.lyzr.com/blog/restrict-amazon-bedrock-models-with-scps.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
