Amazon Nova Micro

I have created an API key on Amazon Bedrock, with the policies AmazonBedrockFullAccess, and AmazonBedrockLimitedAccess attached. Yet when I try to add the key to OpenRouter's BYOK, it says "Key validation failed: Operation not allowed (Tested with: Amazon Bedrock | amazon/nova-micro-v1)"

I have even tested the access on IAM Policy Simulator provided by AWS and both InvokeModel and InvokeModelWithResponseStream returned "allowed" for all resources (\*).

Would greatly appreciate any help... ><

https://preview.redd.it/c089vxbcbrzf1.png?width=382&format=png&auto=webp&s=6e15e50038330a9e881f12de9ebb5b7c605c0afd

https://preview.redd.it/8kal1yvcbrzf1.png?width=1311&format=png&auto=webp&s=96333dc295dc61d8e52bc287c34f6417fa120263

Hey everyone,

I just ran into an issue trying to call **Claude Sonnet 4.5** via the **AWS Bedrock Runtime API**, and I’m hoping someone here might have insights or has faced the same thing.

**Setup:**

* **Account type:** Channel program account (via AWS Partner / Distributor)
* **Region:** `us-east-1`
* **API key:** Valid — works fine for `amazon.nova-micro-v1:0`
* **Model I’m calling:** `anthropic.claude-sonnet-4-5-20250929-v1:0`

Here’s the cURL command I used:

curl -X POST "https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-sonnet-4-5-20250929-v1:0/converse" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <valid-token>" \
-d '{
"messages": [
{
"role": "user",
"content": [{"text": "Hello"}]
}
]
}'

And here’s the **error response** I got back:

{
"message": "Invocation of model ID anthropic.claude-sonnet-4-5-20250929-v1:0 with on-demand throughput isn't supported. Retry your request with the ID or ARN of an inference profile that contains this model."
}

After reaching out to AWS Support, I also got this message:

>

Has anyone here successfully accessed Claude Sonnet 4.5 under a channel program account, or know how to obtain the required inference profile ARN?

I seem i can't use any claude variant of models but I can use aws nova variant tho

Any clarification or workaround would be super appreciated 🙏

Here’s a slightly refined and Reddit-ready version of your post — same message, just cleaner formatting and tone so it reads smoothly and attracts good replies:

# [Help] Can't Access Claude Sonnet 4.5 on AWS Bedrock (Channel Program Account)

Hey everyone,

I just ran into an issue trying to call Claude Sonnet 4.5 via the AWS Bedrock Runtime API, and I’m hoping someone here might have insights or has faced the same thing.

Setup

* Account type: Channel program account (via AWS Partner / Distributor)
* Region: us-east-1
* API key: Valid — works fine for amazon.nova-micro-v1:0
* Model I’m calling: anthropic.claude-sonnet-4-5-20250929-v1:0

cURL command:

curl -X POST "https://bedrock-runtime.us-east-1.amazonaws.com/model/anthropic.claude-sonnet-4-5-20250929-v1:0/converse" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer <valid-token>" \
-d '{
"messages": [
{
"role": "user",
"content": [{"text": "Hello"}]
}
]
}'

Error response:

{
"message": "Invocation of model ID anthropic.claude-sonnet-4-5-20250929-v1:0 with on-demand throughput isn't supported. Retry your request with the ID or ARN of an inference profile that contains this model."
}

After reaching out to AWS Support, I got this message back:

>

It seems like I can’t use any Claude variant (Sonnet, Haiku, etc.), but I can use AWS Nova models just fine.

Has anyone here successfully accessed Claude Sonnet 4.5 under a channel program account, or know how to obtain the required inference profile ARN?

Any clarification or workaround would be super appreciated 🙏

Today, the AI model that hallucinates the least is Google Gemini 2.0 Flash 001, with a factual consistency rate of 99.3%. This score is encouraging because it means that we're relatively close to solving the hallucination problem.

https://github.com/vectara/hallucination-leaderboard

What would happen if we built an AI agent that would first query Google Gemini 2.5 Pro about something, (because it is currently the most powerful model, completely dominating the Chatbot Arena Leaderboard by almost 40 points) and then ran the answer it generated by other models to catch any inaccuracies it may have generated?

https://lmarena.ai/?leaderboard

We presume that the different AI developers use different data sets to build their models, so while one may hallucinate about a certain query, it's possible that another would not. What would happen if we instructed our AI agent to run the content Gemini 2.5 generated through the next ten models by other developers, asking them each to analyze the answer for factual consistency?

Could this be a way to arrive at a factual consistency for answers that is perhaps 99.9% or higher? Could this be done relatively inexpensively and completely automatically?

Below are ten top models that our AI agent would run Gemini 2.5's answer through, ranked according to their factual consistency rate score. I asked 2.5 to comment on the usefulness of this approach, and point out some possible flaws and limitations, and end this post with its answer. While the technique may be relatively expensive, in some cases the usefulness of the accuracy may justify the extra expense.

OpenAI-o3-mini-high-reasoning - 99.2%
Zhipu AI GLM-4-9B-Chat - 98.7%
Amazon Nova-Micro-V1 - 98.4%
Qwen2-72B-Instruct - 95.3%
Mixtral-8x22B-Instruct-v0.1 - 95.3%
Anthropic Claude-3-5-haiku - 95.1%
01-AI Yi-1.5-9B-Chat - 95.1%
Cohere Command-R - 95.1%
IBM Granite-3.0-8B-Instruct - 93.5%
Microsoft phi-2 - 93.3%

Gemini 2.5 Pro:

"Okay, here's a concise summary of the analysis:
Your proposal involves using Gemini 2.5 Pro to generate an answer and then having 10 other models (ranked high for factual consistency) verify it. The goal is automated, inexpensive, near-perfect (99.9%+) factual accuracy.
Potential & Feasibility:
* Improved Accuracy: Yes, this ensemble approach could significantly increase factual accuracy compared to using a single model by catching diverse errors. Reaching 99.9%+ is ambitious but directionally plausible.
* Automation: The process is fully automatable using APIs.
Key Challenges & Limitations:
* Cost & Speed: Running 11 models per query will be substantially more expensive and slower than using one. "Relatively inexpensive" is unlikely.
* Consensus Logic: Deciding how to handle disagreements among the 10 verifiers is complex. A simple majority vote might miss valid corrections or reinforce errors.
* Verifier Imperfection: The verification models themselves aren't perfect and can make mistakes (hallucinate) or fail to spot errors.
* Shared Errors: Models might share common misconceptions from similar training data, leading to false consensus.
* Complexity: Designing effective verification prompts and consensus mechanisms is non-trivial.
In Short:
The idea leverages model diversity for better accuracy and is automatable. However, it faces significant hurdles in cost, speed, and the complexity of managing verification results. While promising, it's not a simple or cheap solution for achieving near-perfect factual consistency."

Today, the AI model that hallucinates the least is Google Gemini 2.0 Flash 001, with a factual consistency rate of 99.3%. This score is encouraging because it means that we're relatively close to solving the hallucination problem.

https://github.com/vectara/hallucination-leaderboard

What would happen if we built an AI agent that would first query Google Gemini 2.5 Pro about something, (because it is currently the most powerful model, completely dominating the Chatbot Arena Leaderboard by almost 40 points) and then ran the answer it generated by other models to catch any inaccuracies it may have generated?

https://lmarena.ai/?leaderboard

We presume that the different AI developers use different data sets to build their models, so while one may hallucinate about a certain query, it's possible that another would not. What would happen if we instructed our AI agent to run the content Gemini 2.5 generated through the next ten models by other developers, asking them each to analyze the answer for factual consistency?

Could this be a way to arrive at a factual consistency for answers that is perhaps 99.9% or higher? Could this be done relatively inexpensively and completely automatically?

Below are ten top models that our AI agent would run Gemini 2.5's answer through, ranked according to their factual consistency rate score. I asked 2.5 to comment on the usefulness of this approach, and point out some possible flaws and limitations, and end this post with its answer. While the technique may be relatively expensive, in some cases the usefulness of the accuracy may justify the extra expense.

OpenAI-o3-mini-high-reasoning - 99.2%
Zhipu AI GLM-4-9B-Chat - 98.7%
Amazon Nova-Micro-V1 - 98.4%
Qwen2-72B-Instruct - 95.3%
Mixtral-8x22B-Instruct-v0.1 - 95.3%
Anthropic Claude-3-5-haiku - 95.1%
01-AI Yi-1.5-9B-Chat - 95.1%
Cohere Command-R - 95.1%
IBM Granite-3.0-8B-Instruct - 93.5%
Microsoft phi-2 - 93.3%

Gemini 2.5 Pro:

"Okay, here's a concise summary of the analysis:
Your proposal involves using Gemini 2.5 Pro to generate an answer and then having 10 other models (ranked high for factual consistency) verify it. The goal is automated, inexpensive, near-perfect (99.9%+) factual accuracy.
Potential & Feasibility:
* Improved Accuracy: Yes, this ensemble approach could significantly increase factual accuracy compared to using a single model by catching diverse errors. Reaching 99.9%+ is ambitious but directionally plausible.
* Automation: The process is fully automatable using APIs.
Key Challenges & Limitations:
* Cost & Speed: Running 11 models per query will be substantially more expensive and slower than using one. "Relatively inexpensive" is unlikely.
* Consensus Logic: Deciding how to handle disagreements among the 10 verifiers is complex. A simple majority vote might miss valid corrections or reinforce errors.
* Verifier Imperfection: The verification models themselves aren't perfect and can make mistakes (hallucinate) or fail to spot errors.
* Shared Errors: Models might share common misconceptions from similar training data, leading to false consensus.
* Complexity: Designing effective verification prompts and consensus mechanisms is non-trivial.
In Short:
The idea leverages model diversity for better accuracy and is automatable. However, it faces significant hurdles in cost, speed, and the complexity of managing verification results. While promising, it's not a simple or cheap solution for achieving near-perfect factual consistency."

Today, the AI model that hallucinates the least is Google Gemini 2.0 Flash 001, with a factual consistency rate of 99.3%. This score is encouraging because it means that we're relatively close to solving the hallucination problem.

https://github.com/vectara/hallucination-leaderboard

What would happen if we built an AI agent that would first query Google Gemini 2.5 Pro about something, (because it is currently the most powerful model, completely dominating the Chatbot Arena Leaderboard by almost 40 points) and then ran the answer it generated by other models to catch any inaccuracies it may have generated?

https://lmarena.ai/?leaderboard

We presume that the different AI developers use different data sets to build their models, so while one may hallucinate about a certain query, it's possible that another would not. What would happen if we instructed our AI agent to run the content Gemini 2.5 generated through the next ten models by other developers, asking them each to analyze the answer for factual consistency?

Could this be a way to arrive at a factual consistency for answers that is perhaps 99.9% or higher? Could this be done relatively inexpensively and completely automatically?

Below are ten top models that our AI agent would run Gemini 2.5's answer through, ranked according to their factual consistency rate score. I asked 2.5 to comment on the usefulness of this approach, and point out some possible flaws and limitations, and end this post with its answer. While the technique may be relatively expensive, in some cases the usefulness of the accuracy may justify the extra expense.

OpenAI-o3-mini-high-reasoning - 99.2%
Zhipu AI GLM-4-9B-Chat - 98.7%
Amazon Nova-Micro-V1 - 98.4%
Qwen2-72B-Instruct - 95.3%
Mixtral-8x22B-Instruct-v0.1 - 95.3%
Anthropic Claude-3-5-haiku - 95.1%
01-AI Yi-1.5-9B-Chat - 95.1%
Cohere Command-R - 95.1%
IBM Granite-3.0-8B-Instruct - 93.5%
Microsoft phi-2 - 93.3%

Gemini 2.5 Pro:

"Okay, here's a concise summary of the analysis:
Your proposal involves using Gemini 2.5 Pro to generate an answer and then having 10 other models (ranked high for factual consistency) verify it. The goal is automated, inexpensive, near-perfect (99.9%+) factual accuracy.
Potential & Feasibility:
* Improved Accuracy: Yes, this ensemble approach could significantly increase factual accuracy compared to using a single model by catching diverse errors. Reaching 99.9%+ is ambitious but directionally plausible.
* Automation: The process is fully automatable using APIs.
Key Challenges & Limitations:
* Cost & Speed: Running 11 models per query will be substantially more expensive and slower than using one. "Relatively inexpensive" is unlikely.
* Consensus Logic: Deciding how to handle disagreements among the 10 verifiers is complex. A simple majority vote might miss valid corrections or reinforce errors.
* Verifier Imperfection: The verification models themselves aren't perfect and can make mistakes (hallucinate) or fail to spot errors.
* Shared Errors: Models might share common misconceptions from similar training data, leading to false consensus.
* Complexity: Designing effective verification prompts and consensus mechanisms is non-trivial.
In Short:
The idea leverages model diversity for better accuracy and is automatable. However, it faces significant hurdles in cost, speed, and the complexity of managing verification results. While promising, it's not a simple or cheap solution for achieving near-perfect factual consistency."