Challenge of the Day

Making Sense of Bedrock Model Billing and the Attack of the AWS Marketplace Agreements You Are Receiving

Marketplace Central team · November 10, 2026 · 6 min read

As your company begins to consume more and more AI models via Amazon Bedrock, you will be trying to make sense of the billing and the AWS Marketplace agreements that are being created. At the end of the month you will try to understand which agreement was for which billing, and what the rate of consumption was. These agreements are created by usage, not by procurement. A developer in a sandbox account calling a model is enough. They often arrive as $0 agreements that simply set the consumption rate, much like a free trial, with no purchase order, no owner, and no procurement review. Too often you find out at invoice time, after the commitment already exists.

What is happening

Every time a third-party model is invoked for the first time in an AWS account, Amazon Bedrock creates an agreement for it to be billed against. That is per model and per account, not once per vendor or once per company. AWS documents this, and AWS's own list of Marketplace-sold models includes Anthropic Claude. OpenAI models behave the same way: when a team makes their first call to an OpenAI model in Bedrock, it produces an AWS Marketplace agreement, just as Anthropic models do. Your token rate is billed at the rate of that agreement.

One first call, one agreement
Anthropic model
OpenAI model
Another third-party model
Sandbox account
Agreement
Agreement
Not used yet
Data science account
Agreement
Not used yet
Not used yet
Production account
Agreement
Agreement
Agreement
  • Created by usage, not procurement
  • Often $0 at creation
  • No purchase order
  • No named owner

Illustrative. Bedrock creates an agreement per model and per AWS account, not once per vendor or once per company.

What is the rate? That depends, but most likely standard rates, unless you have a separate agreement you've negotiated. If you have negotiated a rate, it is likely not vendor-wide. It is model-specific and for a specific AWS account, and it does not follow the workload on its own. When a new account is not covered, it quietly creates a standard agreement and the discount leaks without anyone noticing.

The discount does not follow the workload

Account A

Named in the negotiated agreement

Negotiated rate

Account B

New account, not covered

New standard agreement, standard rate

Account A, new model

Model not named in the agreement

New standard agreement, standard rate

Illustrative. A negotiated rate is usually model-specific and tied to a specific AWS account. When usage lands somewhere it does not cover, a standard agreement is created and the discount leaks without anyone noticing.

So you can get a new AWS Marketplace agreement for a model you thought was covered. A team starts using it in an account your agreement does not cover, or it turns out to be a different model than the one your agreement covers. For example, an agreement built from today's Anthropic or OpenAI bill that names Sonnet may not cover the next model your teams adopt. Services that call models on your behalf, such as AgentCore, may add to this, and we are still verifying how those agreements show up. As more and more workloads move to Amazon Bedrock-based inference, you are going to come under more of these agreements.

OpenAI also reached AWS Marketplace a second way. On November 9, 2026, ChatGPT Enterprise was listed as a SaaS product bought through a private offer, with spend counting toward your AWS commitment. That is a deliberate purchase, separate from the usage-driven Bedrock agreements, and it adds one more agreement that someone has to own.

What can you do to be proactive?

Right now it's almost impossible to do anything proactively without being a blocker to AI modernization within your organization. What needs to be done immediately is to understand which models are running, in which accounts, and what your growth plans are. There are no guarantees that any pricing discounts will be available, but without the data you won't be able to put together your plan for your discussion with AWS or your model vendor.

Ask your cloud team

  • 1Which AWS accounts have created agreements for models, and which sellers?
  • 2Which of those agreements have a named owner and a purchase order, and which have neither?
  • 3Is our negotiated agreement shared with every account that uses the model, or are some accounts paying standard rates?
  • 4Does our negotiated agreement name every model our teams are likely to adopt next, or only the ones on today's bill?

Can one agreement cover more accounts?

For a negotiated private offer on some Bedrock models, yes. AWS calls it managed entitlements. The account that accepted the offer shares it with other accounts, an organizational unit, or the whole AWS Organization, and once the shared access is activated, they run on the negotiated rate. Your cloud team sets it up, and AWS's October post, Manage private pricing offers for Amazon Bedrock with managed entitlements, walks them through it. AWS's earlier June post, Simplify multi-account access to Amazon Bedrock models with managed entitlements, covers what happens when an account already has its own subscription. The risk for you is that someone has to remember to do it for every new account. You should only have to say, "this account should come under that agreement."

How one agreement reaches more accounts
  1. Step 1

    Accept

    One account accepts the negotiated private offer.

  2. Step 2

    Share

    It is shared with other accounts, an organizational unit, or the whole AWS Organization.

  3. Step 3

    Activate

    The shared access becomes active in the account.

  4. Step 4

    Negotiated rate

    Usage there runs on the negotiated rate.

Until the shared access is active, new usage in that account runs on standard rates.

AWS calls this managed entitlements. Your cloud team sets it up, and someone has to remember to do it for every new account.

Three catches

  1. Catch 1First, it covers Bedrock models backed by AWS License Manager, not SaaS products.
  2. Catch 2Second, the shared access has to be active before the negotiated rate applies, so until then new usage there runs on standard rates.
  3. Catch 3Third, if an account already created its own agreement, AWS says the entitlement from that first agreement is disabled once the shared access is active, which leaves the first agreement for you to review and close out.

Can you add a Purchase Order to AI models?

Once you have an agreement for your model consumption across multiple accounts, it makes sense to add a purchase order to the consumption. Like all AWS Marketplace agreements, purchase orders can be added to agreements via the Agreement API. Without one, invoices cannot be matched to a PO in your ERP or procure-to-pay system.

What do you see on the invoice?

Like other AWS Marketplace purchases, they are shown separately on your bill by product name with the monthly total. As an example, if teams have started using Claude Sonnet 4.6 in 3 accounts that create 3 agreements, you will see a rollup of spending for Claude Sonnet 4.6, not the account breakdown or the agreements that have filled your inbox. Additionally, the unit prices for 1M tokens (input/output) are not on the bill, just the number of tokens and the monthly fee total. That is why procurement typically finds out at invoice time and has little to reconcile against.

Three agreements, one line on the bill
Account 1Claude Sonnet 4.6 agreement
Account 2Claude Sonnet 4.6 agreement
Account 3Claude Sonnet 4.6 agreement

On the invoice

Claude Sonnet 4.6Tokens used, monthly total
  • Not shown: Breakdown by account
  • Not shown: Which agreement it belongs to
  • Not shown: Unit price per 1M input and output tokens

From the example in the post: Claude Sonnet 4.6 used in 3 accounts creates 3 agreements. The invoice rolls them up by product name.

What we think is needed

As Amazon Bedrock model consumption increases, we need to aggregate and organize the agreements that are in place to validate whether the rates being charged are standard or negotiated. Organizing them by vendor, with each purchase order attached to the right agreement, gives procurement one view of what is governing spend. Without knowing which agreements are governing pricing in which accounts, we can't really start negotiations with AWS, nor actually govern the pricing rate per model. Once we have agreements aggregated, we need to add purchase orders to usage like any other AWS Marketplace purchase and extend negotiated agreements to new accounts so the discount follows the workload. We then need to automate the reconciliation process to our accounts payable systems so that we accrue billing correctly. We also need to know whether that consumption counts toward our existing commitments, such as an EDP or PPA, so AI spend is not running outside them. All of this needs to use AI to automate, and have a simple UI that provides guidance on how to govern and manage Amazon Bedrock agreements and billing.

What is needed
  1. 1Aggregate and organize
  2. 2Attach purchase orders
  3. 3Extend to new accounts
  4. 4Reconcile to AP

Also know whether this consumption counts toward your existing commitments, such as an EDP or PPA.

Automated with AI, with a simple UI that guides how to govern and manage Amazon Bedrock agreements and billing.

More Challenge of the Day posts