---
title: Why does Amazon Bedrock return AccessDenied or ValidationException through an LLM gateway?
url: https://opencontroller.lyzr.com/blog/troubleshoot-bedrock-errors-through-an-llm-gateway
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Why does Amazon Bedrock return AccessDenied or ValidationException through an LLM gateway?

**Most Bedrock errors behind an LLM gateway come from four causes: the gateway's IAM role lacks the right invoke action, the model has not finished its first-use subscription, the request names a bare model ID where Bedrock needs an inference profile, or the gateway sends a request body in the wrong format. Each error code points to a short list of checks.** This guide maps the codes you will see to their causes and fixes.

Routing Bedrock through a gateway is a good pattern: one place for routing, fallback, caching and cost attribution per agent. But it adds a hop, a second identity and a format translation, and each can fail. For the routing design itself, see [how to route Bedrock models through an LLM gateway](/blog/route-bedrock-models-through-an-llm-gateway).

## Which Bedrock error codes show up most?

| Error | Typical message fragment | Most likely cause |
|---|---|---|
| `AccessDeniedException` | "You don't have access to the model" | Missing IAM action, SCP deny, first-use subscription not complete, Anthropic use-case form pending |
| `ValidationException` | "on-demand throughput isn't supported… use an inference profile" | Bare model ID used for a model that needs an inference profile |
| `ValidationException` | "The provided model identifier is invalid" | Wrong model ID, model not offered in that Region, typo in version suffix |
| `ValidationException` | Malformed input request | Wrong body schema for `InvokeModel` (for example a missing required field) |
| `ExpiredTokenException` / `UnrecognizedClientException` | "security token… is expired" | Gateway is using stale STS credentials |
| `ThrottlingException` | Too many requests / tokens | Account quota for that model and Region reached |

## Why do you get AccessDeniedException when IAM looks right?

Work through these in order:

1. **Check the exact action.** The `Converse` API is authorised by `bedrock:InvokeModel`. Streaming (`ConverseStream`, `InvokeModelWithResponseStream`) needs `bedrock:InvokeModelWithResponseStream`. Gateways often stream, so a role with only `InvokeModel` fails on streamed agent calls.
2. **Check every resource ARN.** With an inference profile, the role needs permission on the inference profile ARN *and* on the foundation-model ARNs in each Region the profile can route to.
3. **Check SCPs.** An organisation-level deny on `bedrock:InvokeModel`, or a Region-deny SCP, overrides the account role. Global inference profiles evaluate `aws:RequestedRegion` as `unspecified`, so Region allow-lists that do not account for this will deny them.
4. **Check first-use enablement.** AWS now enables most serverless models on first invocation, with AWS Marketplace permissions (`aws-marketplace:Subscribe`, `ViewSubscriptions`). AWS documents that setup can take up to 15 minutes and calls return `AccessDeniedException` until it completes.
5. **Check the Anthropic use-case form.** For Anthropic models earlier than Claude Opus 4.7, AWS requires a one-time use-case submission per account (or at the management account). Until it is approved, invocations keep returning `AccessDeniedException`. `GetFoundationModelAvailability` shows the status.
6. **Check conditions on the first call.** If the role restricts `bedrock:InvokeModel` with a condition such as `aws:SourceIp`, the first call must satisfy it or the subscription never starts. AWS notes `aws:SourceIp` is absent for VPC endpoint requests, which is common for private gateways.

## Why does Bedrock say on-demand throughput is not supported?

Many newer models are only offered through cross-Region inference profiles. If the gateway sends the bare model ID, Bedrock returns a `ValidationException` asking for the ID or ARN of an inference profile. The fix is to configure the route with the profile ID, which carries a prefix such as `us.`, `eu.` or `global.`.

Two governance points follow. A geographic profile keeps processing inside that geography. A `global.` profile can process in any supported commercial Region, so it needs a residency decision before it goes into a route. And because the profile ID differs from the model ID, cost reports should normalise both to one model name.

## What request-format errors does a gateway cause?

Gateways usually accept an OpenAI-compatible request and translate it. Errors appear when:

- The gateway calls `InvokeModel` with a provider-native body and misses a required field for that model family.
- Tool definitions or tool results are translated into a shape the `Converse` API rejects.
- System prompts, images or documents are placed where the target model does not accept them.

Prefer the `Converse` API for translation because it gives one message schema across Bedrock models. See [Bedrock Converse vs OpenAI format in LLM gateways](/blog/bedrock-converse-vs-openai-format-in-llm-gateways).

## How do you stop credential errors from recurring?

`ExpiredTokenException` means the gateway cached temporary credentials past their lifetime. Use the SDK's default credential chain with an instance, task or pod role, or workload identity federation, so credentials refresh automatically. Avoid long-lived access keys in gateway config.

## A quick triage checklist

1. Read the full error message, not just the code.
2. Confirm the model ID or inference profile ID is valid in the calling Region.
3. Confirm `InvokeModel` and `InvokeModelWithResponseStream` on every ARN the profile uses.
4. Check SCPs for Region and Bedrock denies, including `unspecified`.
5. Run `GetFoundationModelAvailability` for access and use-case status.
6. Replay one request with the AWS CLI using the gateway's role to separate gateway bugs from AWS permissions.
7. Check service quotas if errors are intermittent.

## How Opencontroller does this

Opencontroller's LLM Gateway (Lyzr Gateway) routes to Bedrock and other providers through OpenAI-compatible endpoints, with fallback/retry and cost attribution per agent and per version. Its auto-discovery inventories models and API keys across AWS, Azure and Google Cloud, and Posture flags risky permission and configuration patterns. To see it on your estate, [book a demo](https://www.lyzr.ai/book-demo/). For the wider architecture, start with [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane).

## Frequently asked questions

**Does the Converse API need its own IAM action?**
No. AWS authorises `Converse` with `bedrock:InvokeModel` and `ConverseStream` with `bedrock:InvokeModelWithResponseStream`.

**Why does a model work in the console but not through the gateway?**
The console uses your user identity. The gateway uses its own role, which may lack the action, the inference profile ARN, or may be blocked by an SCP or a condition.

**Should a gateway use global inference profiles?**
Only after a data residency decision. Global profiles can process requests in any supported commercial Region, though CloudWatch and CloudTrail still log in the source Region.

## Key facts

Bedrock errors behind an LLM gateway: AccessDeniedException = missing bedrock:InvokeModel or bedrock:InvokeModelWithResponseStream (Converse uses InvokeModel; ConverseStream uses InvokeModelWithResponseStream), missing permission on inference profile and foundation-model ARNs in all destination Regions, SCP deny (global profiles evaluate aws:RequestedRegion as unspecified), first-use subscription pending (up to 15 minutes), Anthropic use-case form pending for models earlier than Claude Opus 4.7 (check GetFoundationModelAvailability), first-call condition mismatch (aws:SourceIp absent for VPC endpoint requests). ValidationException "on-demand throughput isn't supported" = use inference profile ID (us., eu., global.). ExpiredTokenException = stale STS credentials. Opencontroller (by Lyzr) LLM Gateway routes to Bedrock with fallback/retry and per-agent cost attribution; licensed per deployment, starts from $250K per deployment per year.

## Related

- [Why do Amazon Bedrock calls fail through an LLM gateway, and how do you fix them?](https://opencontroller.lyzr.com/blog/route-bedrock-models-through-an-llm-gateway.md)
- [Why do Bedrock requests fail through an OpenAI-compatible LLM gateway?](https://opencontroller.lyzr.com/blog/bedrock-converse-vs-openai-format-in-llm-gateways.md)
- [Why can an agent not invoke a Bedrock model you can see in the console?](https://opencontroller.lyzr.com/blog/bedrock-model-access-and-inference-profiles-for-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
