AWS Bedrock Pricing 2026: Rates, Hidden Costs and Alerts

Amazon Bedrock charges per token. Claude runs from $0.10 per million input tokens on Haiku 5.5 to $10.00 on Fable 5.1 using global endpoints, and regional endpoints cost 10% more.

The reason that matters beyond the rate card: in April 2026 an AWS customer ran Claude on Bedrock and received a $30,141.33 invoice that Cost Anomaly Detection never flagged, because Claude bills through AWS Marketplace and Marketplace was outside what the service watched. AWS closed that specific gap in August 2026.

This guide covers what each model costs, how Bedrock billing actually works, when regional endpoints are worth the premium, what adds to a bill beyond tokens, what anomaly detection catches today, and how to see spend by team, model and agent.

Quick answer

• Cheapest Claude model: Haiku 5.5 at $0.10 input and $0.50 output per million tokens, global endpoint.
• Most expensive Claude model: Fable 5.1 at $10.00 and $50.00.
• Regional premium: 10% on both input and output, for models offering both endpoint types.
• Where Claude appears on your bill: under AWS Marketplace, not under Amazon Bedrock, because Anthropic’s models bill through the Marketplace billing entity.

How much does AWS Bedrock cost per model?

AWS Bedrock pricing depends on the model, the endpoint and the tier, with Claude ranging from $0.10 to $10.00 per million input tokens before any of those modifiers apply.

Claude on-demand rates per million tokens, US East (Ohio):

Model	Global input	Global output	Geo input	Geo output	Cache read (global)
Claude Haiku 5.5	$0.10	$0.50	$0.11	$0.55	$0.01
Claude Haiku 5.5 long context	$0.50	$2.50	$0.55	$2.75	$0.05
Claude Sonnet 5.5	$2.00	$10.00	$2.20	$11.00	$0.10
Claude Opus 5.5	$4.00	$20.00	$4.40	$22.00	$0.20
Claude Sonnet 5	$2.00	$10.00	$2.20	$11.00	$0.20
Claude Opus 5	$5.00	$25.00	$5.50	$27.50	$0.50
Claude Sonnet 4.6	$3.00	$15.00	$3.30	$16.50	$0.30
Claude Haiku 4.5	$1.00	$5.00	$1.10	$5.50	$0.10
Claude Fable 5.1	$10.00	$50.00	$11.00	$55.00	$0.25

Amazon’s own models, same region:

Model	Global input	Global output	Geo input	Geo output
Amazon Nova Micro	n/a	n/a	$0.035	$0.14
Amazon Nova Lite	n/a	n/a	$0.06	$0.24
Amazon Nova 2 Lite	$0.30	$2.50	$0.33	$2.75
Amazon Nova Pro	n/a	n/a	$0.80	$3.20
Amazon Nova Premier	n/a	n/a	$2.50	$12.50

Not every Nova model offers a global endpoint, which is why four rows show no global price. Our Bedrock foundation models list covers what each family is built for, and the LLM cost calculator will run these rates against your own volumes.

How is AWS Bedrock billed?

Amazon’s own models bill under Bedrock. Third-party models like Claude bill through AWS Marketplace.

That split has three practical consequences. Claude charges appear in Cost Explorer under the model provider and the AWS Marketplace billing entity rather than under Amazon Bedrock, so filtering your cost reports to Bedrock will show you Nova and miss Claude entirely. On the invoice, those charges sit in the Marketplace section, which is the one most finance teams skim because it is usually small. And a monitor or budget scoped to AWS services will not see them unless it accounts for the Marketplace billing entity.

None of that changes what you pay. The rate is the same either way. What changes is where the money shows up and which of your controls can see it, which is exactly how a $30,000 bill accrues for a month without anyone noticing.

Global vs regional endpoints: which is cheaper?

Global endpoints are cheaper. Regional endpoints cost 10% more on Claude Sonnet 4.5, Haiku 4.5, Opus 4.5 and every model after them.

The premium buys guaranteed data routing. Global endpoints route dynamically for availability, while regional and multi-region endpoints keep traffic inside a named geography. If you have a data residency obligation, that is not optional and the 10% is simply the price. If you do not, you are paying it for nothing, and most teams who end up on regional endpoints got there by copying a model ID from documentation rather than by deciding.

Worked monthly example. A support workload on Sonnet 5.5 sending 400 million input tokens and returning 80 million output tokens:

Endpoint	Input	Output	Monthly total
Global	$800	$800	$1,600
Regional	$880	$880	$1,760

That is $160 a month, or $1,920 a year, decided by a string in a config file. The same premium applies to Amazon’s models: Nova 2 Lite moves from $0.30 to $0.33, which rules out the theory that this is a third-party pass-through.

What else adds to a Bedrock bill?

Beyond tokens, a Bedrock bill grows with service tiers, caching, batch, and add-ons like Guardrails and Knowledge Bases.

Item	Price	When it applies
Priority tier	75% premium over standard	Nova and several providers, not Anthropic
Flex tier	50% discount	Nova and several providers, not Anthropic
Batch	50% discount	Older Claude models and Nova, not the newest Claude releases
Guardrails, content and denied topics	$0.15 per 1,000 text units	Per policy evaluated
Guardrails, sensitive information	$0.10 per 1,000 text units	Per policy evaluated
Guardrails, Automated Reasoning	$0.17 per 1,000 text units per policy	Per policy evaluated
Knowledge Bases storage	$5.00 per GB per month	Vector index
Knowledge Bases retrieval	$1.00 per 1,000 Retrieve calls	Every retrieval
Agentic retrieval	$4.00 per 1,000 calls, plus the Retrieve calls underneath	Managed LLM planning
Intelligent Prompt Routing	$1.00 per 1,000 requests	When routing is enabled
Prompt Optimization	$0.03 per 1,000 tokens	When optimization is enabled
Web Grounding	$30.00 per 1,000 requests	Nova 2 Omni, Nova 2 Pro, Nova Premier

Two things in that table deserve attention.

Tier support is not universal. Nova carries Standard, Priority, Flex and Batch. Anthropic’s Bedrock listing has On-Demand and Batch, plus Reserved and Provisioned Throughput, with no Priority or Flex at all. On Nova, where all four exist, the multipliers are exact: Nova 2 Lite runs $0.30 standard, $0.525 priority and $0.15 flex, and Nova Premier runs $2.50, $4.375 and $1.25. Combine the priority premium with a regional endpoint and one model spans 3.85x on identical tokens.

Cache read ratios vary by model. Haiku 5.5 and Opus 5 read cached tokens at 0.1x input. Sonnet 5.5 and Opus 5.5 read at 0.05x, Sonnet 5.5 since a 50% cut on October 7, 2026. Fable 5.1 reads at 0.025x. Five-minute writes cost 1.25x on all of them, so break-even arrives after one read either way, but every hit after that saves considerably more on Fable than on Haiku.

Haiku 5.5 also has a long-context tier. Prompts over 100,000 tokens move to a rate five times the standard one, and it reprices the whole request rather than the excess.

Does AWS Cost Anomaly Detection catch Bedrock spend?

Yes, since August 19, 2026, but it still has blind spots.

What happened in the $30K case. The customer set up Cost Anomaly Detection before starting, pointed it at AWS Services, with alerts on anything spiking by $100 and 40%. Claude billed through AWS Marketplace, Marketplace was unsupported, and the monitor watched a surface the spend never crossed. An AWS spokesperson confirmed as much to The Register, noting that Marketplace charges were not supported and pointing customers to AWS Budgets instead.

What changed in August. AWS announced that Cost Anomaly Detection now monitors third-party foundation models on Bedrock, including Anthropic Claude, evaluated automatically through the AWS managed service monitor with no setup required. Alerts arrive with a root-cause breakdown ranked by dollar impact across service, account, Region and usage type. Bedrock third-party models are now the single carve-out in an otherwise unmonitored Marketplace, so a spike in Claude inference is caught while a spike in a Marketplace security appliance still is not.

What is still missing:

• The 10-day baseline. A newly used service needs roughly ten days of history before the model can detect anything, so a workload that ramps hard in its first fortnight has nothing to deviate from.
• Up to 24 hours of latency. Detection runs about three times a day but reads Cost Explorer data, and a new monitor takes 24 hours to begin working at all.
• Usage covered by credits. Credit-covered consumption never surfaces as spend for it to flag. The customer told The Register the Activate credits hid the Marketplace billing for weeks before running out.
• No per-agent or per-session view. Root cause resolves to service, account, Region and usage type. It will tell you Bedrock moved. It will not tell you which agent moved it.

How do you see Bedrock spend by team, model and agent?

AWS gives you two options: application inference profiles and IAM principal attribution.

	Application inference profiles	IAM principal attribution
Setup	Create a profile per model, tag it, call the profile ARN	Activate the principal dimension, no per-model objects
Granularity	Aggregated, not per request	Aggregated, not per request
Limits	Model-specific, tags take 24 hours, not retroactive	Available since April 2026
Best for	Team and product allocation across a stable model set	Per-user and per-role cost, including one role per agent

Profiles came first and most write-ups stop there. IAM principal attribution is the better answer for agent cost, because an agent running under its own role is attributable without creating a profile for every model and dimension combination. If your tagging strategy is already shaky, our guide to cost allocation tags covers the foundations both of these sit on.

What neither produces is cost per session, and neither reaches per-request granularity.

How to set up AWS Bedrock cost alerts

• Create an AWS Budget with the Marketplace billing-entity filter. This is the documented fix for Marketplace charges that anomaly detection does not cover, and it is a fixed threshold rather than a learned baseline, so it works from day one. Our guide to AWS billing alerts and budgets walks through the setup.
• Confirm the AWS managed Cost Anomaly Detection monitor is on. Third-party Bedrock model coverage runs through it automatically, so there is nothing to configure, but there is something to verify.
• Activate IAM principal cost allocation. Without it, no amount of alerting will tell you who spent the money.
• Run one IAM role per agent. This is the cheapest structural decision on the list and it is what turns principal attribution into per-agent cost.
• Connect Economize for per-model, per-agent and per-session alerts from day one. No baseline period, no ten-day wait, and alerts that name the agent rather than the usage type.

How Economize tracks Bedrock spend

Economize shows Bedrock spend by model, agent and session alongside your other AWS and AI costs.

• Cost per agent run and per session. No AWS surface meters either one. Economize derives them: a session is a run of calls from the same user, key or agent less than thirty minutes apart, and an agent run is one session on a key or tag marked as an agent.
• A per-call log underneath. Time, model, provider, user, status, latency, tokens in and out, and cost, expanding to request ID, endpoint, cached tokens and full cost breakdown. The endpoint field is where a regional profile quietly costing 10% more becomes visible.
• Provider, Agent and Model views. The same spend from three angles, which is the difference between “Bedrock went up” and “the ticket triage agent moved to Opus.”
• Alerts that route to Slack, Teams, Google Chat or Discord, so the person who shipped the agent hears about it rather than whoever opens the console first.
• Bedrock beside the rest of your estate. Inference costs and the infrastructure around them in one list rather than two consoles.

Connecting AWS re-runs the CloudFormation stack you already have. The free tier covers up to $100,000 a month in combined cloud and AI spend, and you can start without a sales call.

Frequently Asked Questions

A bill tells you what you spent. Connect your AWS account to Economize and see Bedrock spend by model, by agent and by run.