OpenRouter or AWS Bedrock: where a researcher should buy model access

Both services sell you access to other companies’ models through one account. They are built for different buyers. Terms and prices are as of 9 October 2026.

OpenRouter for public data, Bedrock for licensed data

Use OpenRouter to compare models and to prototype with public data. Use Bedrock when prompts will contain licensed or sensitive data, or when your university already has an AWS agreement and credits.

	OpenRouter	AWS Bedrock
What it is	One key and one bill for many labs’ models	Model access inside your AWS account
Price	The provider’s list price, plus a fee when you buy credits	Per token by model. Batch jobs 50% lower
Who can see your prompts	Not logged by OpenRouter by default. Each model provider has its own policy	Model providers have no access to prompts or completions
Setup	Minutes	An AWS account and IAM permissions
Best for	Comparing models, public data	Licensed data, batch jobs, AWS credits

OpenRouter

Product. One API key and one bill for models from many labs. The full list is in its model browser.

Price. OpenRouter’s FAQ says it passes through the underlying provider’s price with no markup and charges a fee when you buy credits. Its words: “you’ll always pay the same as the provider’s listed price.” If you bring your own provider keys, there is a monthly allowance with no fee and a percentage fee above it.

Your prompts. Two companies handle them, and each has its own terms.

OpenRouter itself logs request metadata such as timestamps, the model and token counts. The FAQ states that prompts and completions are not logged by default. Logging them is opt-in, in exchange for a 1% discount.
The provider that runs the model has its own policy. OpenRouter’s documentation says each provider handles logging and retention differently and that there are exceptions to “prompts will not be trained on.” An account setting lets you block routing to providers that may train on your data, and there is a zero-data-retention option.

Use it for. A new project where you do not yet know which model you need. Text-as-data work on public filings. Checking whether a result holds across providers, which one recent paper shows you should do.

AWS Bedrock

Product. Model access inside an AWS account. The pricing page lists models from AI21 Labs, Amazon, Anthropic, Cohere, DeepSeek, Google, Meta, Mistral AI, OpenAI, Qwen, xAI and others.

Price. Per-token on-demand prices by model, listed on the pricing page. Batch inference is offered at a 50% lower price than on-demand for supported models, which suits a job that scores a million filings overnight.

Your prompts. AWS’s data protection page says models run in accounts operated by the Bedrock service team and that model providers “don’t have access to Amazon Bedrock logs or to customer prompts and completions.” That is a simpler statement than a per-provider table, and it is the one a university counsel or a data vendor will want to see.

Use it for. Anything that touches licensed data. Work paid for by AWS research credits. Schools where purchasing already runs through AWS, since a new vendor can take longer to approve than the project takes to finish.

Setup. An AWS account, permissions set up by someone who knows IAM, and model access that may need to be requested by region. OpenRouter takes minutes. Bedrock can take an afternoon.

Four questions before you buy

Is any licensed or confidential data going into the prompt? Read the data license first. Many do not allow sending the data to a third party at all. If it is allowed, prefer the service whose terms you can show the vendor, which today is Bedrock.
Do you know which model you need? If not, start on OpenRouter with public data and find out.
Is it a large batch job? Price it on Bedrock batch before you price it anywhere else.
Who pays? A personal card works on OpenRouter. A grant or department usually pays AWS more easily.

You can use both: explore on one, run the licensed-data job on the other, and record which service and model produced each result.