Your Bedrock Line Is Not Your AI Bill: Claude Bills as AWS Marketplace
Claude on Amazon Bedrock is not billed as Amazon Bedrock. AWS sells it as a third-party model through AWS Marketplace, so the charges land under the AWS Marketplace billing entity, one service per model, and a report filtered to "Amazon Bedrock" misses every dollar of it. On one AWS Organization I run, that filter showed 0.4% of the model spend. That classification decides more than the report: general promotional credits exclude Marketplace, and until 19 August 2026 Cost Anomaly Detection ignored it, which is why one AWS user's $30,141.33 of Claude inference charges in April raised no alert. Two more gaps are program rules that apply whatever the billing entity: reseller private pricing covers Bedrock for Amazon Nova and Titan only, and Savings Plans have never covered Bedrock. Here is what AWS's own documents say, the two filters that give you one honest AI number, which discounts and credits actually reach the line, and the order to check them in before you report AI spend or sign a commitment.

Bedrock usage
is how the AWS Support fee calculation counts Marketplace billed Claude, while credits and anomaly alerts treat it as Marketplace
0.4%
of Bedrock model spend the Amazon Bedrock service filter showed on one AWS Organization I run, September to November 2026
$30,141.33
of Claude inference charges one AWS user ran up in April 2026 with no alert, before AWS added Marketplace models to anomaly detection in August
The direct answer: Claude on Bedrock is a Marketplace purchase, and the Bedrock filter cannot see it
AWS sells Anthropic's Claude models on Amazon Bedrock as third-party models, billed through AWS Marketplace. The charges sit on your normal AWS invoice, but under the AWS Marketplace billing entity, and in the tests we cite each model appears as its own service: "Claude Sonnet 5.5 (Amazon Bedrock Edition)", not "Amazon Bedrock". AWS's model card for Claude Opus 5.5 says it in one sentence: charges appear "under the model provider (not under Amazon Bedrock)". The Pricing line on your own model's card tells you which side it sits on. A Cost Explorer view or budget filtered to the Amazon Bedrock service therefore shows the models AWS bills natively and none of your Claude spend.
The classification also decides how some tools and terms treat that spend. General promotional credits exclude Marketplace, and Cost Anomaly Detection ignored Marketplace billed models until AWS added a named exception on 19 August 2026. Two more gaps are program rules that would apply anyway: reseller private pricing names only Amazon Nova and Titan, and Savings Plans have never covered Bedrock, Amazon's own models included. The order of operations is short: split the bill by billing entity first, then report AI spend, then do commitment maths on what is left.
AWS says it in four places, and the pricing page is not one of them
AWS states it in four separate documents. The Bedrock FAQ answers the question directly: customers "will see an AWS Marketplace bill for certain Bedrock serverless models and Bedrock Marketplace models", because those models are sold by third-party providers. The AWS Service Terms, section 50.12.1, apply the AWS Marketplace terms to models sold by third parties and name Amazon Web Services, Inc. as the invoicing party. The Cost Anomaly Detection guide says third-party foundation models on Bedrock "appear in Cost Explorer and on your bill under the AWS Marketplace billing entity". And each model card carries the line quoted above.
Anthropic's own documentation describes Claude in Bedrock as an AWS-native service billed as AWS service usage, and both statements hold. AWS invoices it. Marketplace classifies it. The invoice arrives from the same company either way; the classification is what your cost tools and your discount instruments read.
It is also not every model. AWS lists 11 providers whose models are not sold through Marketplace: Amazon itself, plus DeepSeek, Google, Meta, MiniMax, Mistral AI, Moonshot AI, NVIDIA, Qwen, xAI and Z.AI, along with OpenAI's open-weight models. A hands-on Cost Explorer test published by Classmethod in September 2026 found exactly that split: Claude and OpenAI's GPT models under AWS Marketplace, Grok under AWS. Two teams calling the same Bedrock API can produce spend on opposite sides of the line, depending only on the model they chose.
One naming trap while you are here: Claude Platform on AWS is a different product, operated by Anthropic and also billed through AWS Marketplace. If both are in use, the Marketplace total holds two kinds of Claude spend.
Why the Bedrock number in your report is too small
Group Cost Explorer by Service and Claude does not appear under Amazon Bedrock. In both Classmethod tests each Marketplace sold model appeared as its own service, named for the model with "(Amazon Bedrock Edition)" appended, and a new service name appeared for each new model the team tried. Classmethod's second test, in October 2026, filtered Service to Amazon Bedrock and put the result in a heading: "Filtering by Amazon Bedrock Alone Excludes Claude and GPT". Any saved report, budget or dashboard built on that filter has been understating AI spend since the first Claude call.
AWS shipped the repair on 8 October 2026: a Product attribute dimension in Cost Explorer, Budgets and Cost Management Dashboards that breaks Bedrock cost down by provider, model, inference type and feature, across all those split service names. The data starts on 1 September 2026 and does not reach further back, so for anything earlier the Billing entity filter is the tool.
In the Cost and Usage Report the same spend carries operation names that send people to a search engine. Since 15 January 2026 AWS writes the Bedrock operation type into line_item_operation, where the generic "Usage" label used to be: InvokeModelInference for a standard model call, InvokeModelStreamingInference for a streamed one. Neither is a new service or a surcharge. They are your token usage, labelled by how the application asked for it.
- ✓Find the Marketplace half: Cost Explorer → Filters → Billing entity → AWS Marketplace → Group by Service → last 3 months.
- ✓Get one Bedrock number: Cost Explorer → Group by → Product attribute → Provider (or Model), with no Service filter. Data from 1 September 2026. The blank group collects every cost without a provider value, mostly the rest of your bill, so leave it out of the AI total. Use closed months: on the account in the next section, the open month came back blank.
- ✓The same from the CLI: aws ce get-cost-and-usage --time-period Start=2026-10-01,End=2026-11-01 --granularity MONTHLY --metrics UnblendedCost --group-by Type=PRODUCT_ATTRIBUTE,Key=provider
- ✓Isolate Marketplace from the CLI: swap the group-by for --group-by Type=DIMENSION,Key=SERVICE and add --filter '{"Dimensions":{"Key":"BILLING_ENTITY","Values":["AWS Marketplace"]}}'
- ✓In the CUR: filter bill_billing_entity to AWS Marketplace and group by line_item_operation to see InvokeModelInference against InvokeModelStreamingInference.
What it looked like on a bill I run
I ran these checks on one AWS Organization I run the bill for, covering 1 September to 17 November 2026. The Amazon Bedrock service filter showed 0.4% of the Bedrock model spend. The other 99.6% sat under the AWS Marketplace billing entity in eight services ending "(Amazon Bedrock Edition)", one per Claude model, from Claude Haiku 4.5 to Claude Opus 5.5. A dashboard built on the Bedrock filter would have reported this AI line as a rounding error.
The operation split was just as lopsided: 95% of the Claude spend was InvokeModelStreamingInference and 5% InvokeModelInference. If the streaming operation is the biggest unfamiliar line on your bill, it is ordinary token usage, not a fault.
One finding I did not expect. For September and October, the Product attribute view matched the Billing entity total exactly, with every line tagged Anthropic. For November, the open month, the provider value was blank on every Bedrock line when I checked on 18 November, so the provider view showed none of the month's Claude spend. AWS documents no such lag. It may fill in at month close, and I will recheck in December. Until you have checked your own account mid month, do not make a Product attribute budget your only alert on the models.
The discount perimeter: what actually reaches a Marketplace line
A discount instrument only works on the spend it is defined to cover, and most of them were defined before anyone ran a language model. Walk the list with the Marketplace total from the previous step in hand.
The Enterprise Discount Program is the one instrument with no public answer. We looked for an AWS statement on the enterprise pricing page, in Service Terms section 50.12, in the Bedrock FAQ and on the Bedrock pricing page, and found none. Third-party negotiation guides describe the general Marketplace rule: eligible purchases retire the commitment at full value, typically capped at 25% of the annual commitment, and are billed at the vendor's price, not your EDP rate. Whether Marketplace billed model fees follow that rule, or are treated as Bedrock usage, is contract language. Ask your account team three questions in writing: do these fees receive the EDP discount, do they count toward the commitment, and do they sit inside the Marketplace cap.
If you already hold a private offer, check that it is switched on. AWS's own guidance is blunt about the mechanics: the offer creates a licence in AWS License Manager, grants to member accounts stay disabled until someone activates them, and until then "negotiated pricing does not apply". Accept the offer from the wrong account and member accounts "continue to be billed at public rates". Open AWS License Manager and check that the grants for the offer show as active, not disabled. A discount that was negotiated and never activated is the most expensive kind.
- ✓Savings Plans: no, for any Bedrock model, Amazon's own included. The list of services eligible for Savings Plans names EC2, Lambda, Fargate and SageMaker AI. This gap has nothing to do with Marketplace, but it belongs in the same subtraction.
- ✓Reseller private pricing: Amazon models only. The Solution Provider and Distributor eligible services list, updated 24 September 2026, includes Amazon Bedrock "only with respect to the following models": Amazon Nova and Amazon Titan. It is a named list, so it leaves out Llama and Mistral too, which are not Marketplace billed.
- ✓Promotional credits: not by default. The AWS Promotional Credit Terms list AWS Marketplace as an ineligible service unless AWS authorises it.
- ✓Activate credits: they can. The Activate terms let them offset Marketplace charges for third-party foundation models on Bedrock, subject to the credit's own terms, which is why startups see Claude drawing down credits.
- ✓Marketplace private offer: yes, and it is the intended route. Negotiated rates for third-party Bedrock models arrive as an AWS Marketplace Private Pricing Offer.
The one place AWS counts it as Bedrock usage
There is a single document where the Marketplace label stops applying, and it is the one that calculates a fee. The AWS Support pricing page excludes AWS Marketplace from the charges that Support fees are calculated on, with a footnote: "Bedrock Marketplace and Bedrock serverless purchases are considered Bedrock usage, though you may see a bill through AWS Marketplace." The same page says the fees are calculated on gross AWS charges, before any discounts or credits are applied.
So the same Claude spend counts as Marketplace for promotional credits, has a named exception in anomaly detection, and counts as Bedrock usage for the Support fee. That is a published definition, and it is easy to miss. It does mean a growing Claude line raises the base your Support fee is calculated on, measured before discounts and credits, even in a month when credits paid for the tokens. Put it in the forecast.
The alert that did not exist until August
In May 2026 The Register reported a reader who tried Claude Opus on Bedrock and ran up $30,141.33 of model inference charges in April while relying on Cost Anomaly Detection. No alert fired, because at the time the service did not monitor AWS Marketplace charges at all. The reader's AWS Activate credits, $8,026.54 of them, covered the first weeks; in the reader's words the billing was "silently working for weeks", and the first warning was the invoice.
AWS closed that gap on 19 August 2026. Cost Anomaly Detection now monitors third-party foundation models on Bedrock through the AWS managed service monitor, with no setup required. The exception is narrow, though: the user guide still says the service does not monitor other third-party products sold through AWS Marketplace, and points you to AWS Budgets for those. If your only guardrail predates August, add a budget scoped to the models as well, because a budget is a fixed threshold you set and an anomaly monitor is a model that decides what unusual means.
- ✓Confirm the monitor exists: Billing and Cost Management → Cost Anomaly Detection → Cost monitors → a monitor Managed by AWS on the AWS services dimension, with an alert subscription that reaches a human.
- ✓Add a budget on the models: Budgets → Create budget → Customize (advanced) → Cost budget → filter Product attribute → Provider = Anthropic. AWS supports actual cost alerts only on these budgets, not forecasts. Check it mid month before you trust it: on the account above the provider value was blank for the whole open month.
- ✓For the mid month alert, a second budget filtered on Billing entity = AWS Marketplace. It also counts unrelated subscriptions, which is the point.
- ✓If you run on credits: add a Charge type filter that excludes Credit, so the budget counts usage before credits and the alert fires while the credits are still paying.
The order that keeps the AI number honest
Four steps, in this order. Split the bill by billing entity, because every later number depends on which side of the line the spend sits. Rebuild the AI figure with the Product attribute dimension, so the report counts Claude, GPT and the Amazon models together. Walk the discount perimeter with the Marketplace total in hand, and get the EDP answer in writing. Only then size a commitment, on the base the instrument actually covers. Reverse the order and you commit against spend that no Savings Plan touches, and report a discount on spend that never receives it.
The models will keep changing, and the billing entity is a property of the provider, not of Bedrock. The check takes about ten minutes and is worth repeating whenever a team adopts a new model.
Upload your most recent AWS bill and we will split it by billing entity the way this article does: how much of your AI spend is Marketplace billed, which of your discounts and credits reach it, and what the honest commitment base is.
- For the levers that cut the Claude line itself: capping Bedrock costs with caching and model laddering
- When the AI bill spans Bedrock, OpenAI and Azure: one ledger for LLM spend
- EDP versus Savings Plans, and the sequencing rule
The short version
- • Claude on Amazon Bedrock is sold and billed through AWS Marketplace: AWS's model card says the charges appear "under the model provider (not under Amazon Bedrock)", so a Cost Explorer report filtered to the Amazon Bedrock service leaves it out. On one AWS Organization I run, that filter showed 0.4% of the Bedrock model spend.
- • Not every Bedrock model works this way: AWS lists 11 providers whose models are not sold through Marketplace, Amazon itself plus third parties such as Meta, Mistral AI and xAI.
- • Two filters fix the report: Billing entity isolates the Marketplace lines, and the Product attribute dimension AWS added on 8 October 2026 groups all Bedrock spend by provider and model, with data from 1 September 2026.
- • General promotional credits treat Marketplace as ineligible unless AWS authorises it. Two more gaps are program rules that apply whatever the billing entity: the reseller private pricing list covers Bedrock for Amazon Nova and Titan only, and Savings Plans have never covered Bedrock for any model.
- • EDP treatment is not published, so read your own agreement: ask your account team in writing whether Marketplace billed model fees receive the discount, count toward the commitment, and fall inside the Marketplace cap.
- • Cost Anomaly Detection has covered third-party Bedrock models only since 19 August 2026; before that, one AWS user ran up $30,141.33 of Claude inference charges in April 2026 without an alert.
Frequently Asked Questions
Why is Claude on Bedrock billed as AWS Marketplace?
Because AWS sells Anthropic's models as third-party content through AWS Marketplace. The Bedrock FAQ says customers will see an AWS Marketplace bill for certain Bedrock serverless models, and the model cards say charges appear under the model provider, not under Amazon Bedrock. The charges are still on your AWS invoice, under the AWS Marketplace billing entity, with each model listed as its own service.
Do Savings Plans or an EDP cover Claude on Amazon Bedrock?
Savings Plans do not, for any Bedrock model: the list of eligible services names EC2, Lambda, Fargate and SageMaker AI. EDP treatment is not published. Third-party guides say Marketplace purchases are billed at the vendor's price and typically count toward the commitment up to a cap of 25%, so ask your account team in writing how your agreement treats Marketplace billed model fees. Negotiated Claude pricing arrives as an AWS Marketplace private offer.
What is InvokeModelStreamingInference on my AWS bill?
It is a Bedrock operation type, not a separate service. Since 15 January 2026 AWS writes the operation into the line_item_operation column of the Cost and Usage Report in place of a generic "Usage" label: InvokeModelInference for a standard model call and InvokeModelStreamingInference for a streamed one. Both are ordinary token usage. For Claude and other Marketplace sold models, the lines sit under the AWS Marketplace billing entity.
Do AWS credits cover Claude on Bedrock?
It depends on the credit. AWS Activate credits can offset Marketplace charges for third-party foundation models on Bedrock. General AWS promotional credits list AWS Marketplace as an ineligible service unless AWS authorises it. Check the eligible services on each credit in the Billing console, and remember that credits can hide a growing bill: in the case The Register reported, $8,026.54 of credits ran out before $30,141.33 of charges appeared.
Sources
- AWS docs: Claude Opus 5.5 model card ("billed through AWS Marketplace", "not under Amazon Bedrock")
- AWS: Amazon Bedrock FAQs (why a Marketplace billing entry appears)
- AWS Service Terms, section 50.12 (third-party models and the invoicing party)
- AWS docs: Cost Anomaly Detection (the Marketplace billing entity and the Bedrock exception)
- AWS docs: Request access to models (the 11 providers not sold through Marketplace)
- AWS What's New, 8 October 2026: Bedrock product attributes in Cost Explorer, Budgets and Dashboards
- AWS docs: Cost Explorer filters (Billing entity, and Product attribute data from 1 September 2026)
- AWS What's New, 15 January 2026: Bedrock operation types in cost reports
- AWS What's New, 19 August 2026: Cost Anomaly Detection supports third-party models on Bedrock
- AWS docs: Services eligible for Savings Plans benefits
- AWS: Solution Provider and Distributor eligible services list for private pricing (PDF, updated 24 September 2026)
- AWS Promotional Credit Terms and Conditions (Marketplace as an ineligible service)
- AWS Activate Terms (credits for third-party Bedrock model spend)
- AWS Support pricing (the Bedrock Marketplace footnote in the fee calculation)
- AWS docs: Subscribing to third-party Bedrock models through a private offer
- AWS Machine Learning Blog, 13 October 2026: private pricing offers for Bedrock with managed entitlements
- Anthropic docs: Claude Platform on AWS, and how it differs from Claude in Amazon Bedrock
- Classmethod: billing entity test for Claude, GPT and Grok on Bedrock (September 2026)
- Classmethod: aggregating Bedrock cost by product attribute (October 2026)
- The Register, 14 May 2026: the $30,141.33 Claude invoice
- CloudForecast: AWS EDP and PPA guide (Marketplace rules and the 25% cap)
Hermann Lotter.FinOps practitioner who has led cloud and AI cost optimization inside a 180-person organisation, identifying six-figure annual savings across AWS and LLM spend. He writes Easy Entropy from hands-on engagements, not theory, and takes on scoped AWS FinOps consulting engagements.LinkedIn
Free Assessment
Want this outcome in your AWS bill?
Get a free cloud cost analysis and a prioritized optimization roadmap.
Request Free Analysis →Topics covered
Related Articles
AI Cost Management
The Bedrock Bill Trap: Capping GenAI Costs on AWS in 2026
GenAI spend is forecast to grow 80% in 2026, and most of it lands on the AWS bill before anyone notices. Here is the FinOps playbook for Amazon Bedrock: prompt caching, model laddering, inference profiles, and chargeback that actually works.
Cost Optimization
How to Reduce Your AWS Bill: A 30-Day Step-by-Step Guide
To reduce your AWS bill, work in this order: turn on Cost Explorer and daily CUR to find where the money actually goes, delete idle and orphaned resources, right-size overprovisioned compute and storage, then buy Savings Plans against the baseline that survives. Most accounts that have never been optimized give up 20% to 35% within 30 days, and the first two steps cost nothing and carry no production risk. Buying commitments first is the common mistake: it locks in the waste you have not removed yet.