AWS Bedrock Pricing (2026): Models, Tiers & Real Examples

Updated October 19, 2026 by the CloudForecast team

When building generative AI applications, Amazon Bedrock (or “AWS Bedrock,” as most of us search for it) is one of the most powerful services out there. It’s fully managed, so you can focus on your app while AWS handles the scaling behind the scenes. And it gives you models from Anthropic, OpenAI, Meta, Amazon, Mistral, and others through a single API.

But pricing? That’s where things get complicated. What used to be “pay per token” has grown into six inference options, model-by-model rates, caching discounts, regional routing, and a growing list of add-ons that show up as separate lines on your bill. Some of those lines don’t even say “Bedrock.”

We first wrote this guide in 2024 and rewrote it for 2026 after AWS added Priority, Flex, and Reserved tiers, prompt caching, global cross-Region pricing, and a new free tier. We updated it again in October 2026 for the Claude 5.5 models, Nova 2 Lite, and a few corrections on which models support which tiers.

In this guide, we’ll cover:

• What AWS Bedrock is
• How the six inference options are priced
• What the major models cost per million tokens
• A realistic monthly cost example (and how to cut it)
• How to estimate and track what you’re actually spending
• A few cost optimization strategies
• Frequently asked questions

What is AWS Bedrock?

First things first: Amazon Bedrock is a fully-managed service designed to simplify the development and scaling of generative AI applications. Bedrock provides access to a range of high-performing foundation models (FMs) from top AI companies–all conveniently through a single API.

How Amazon Bedrock pricing works: the six inference options

Every Bedrock bill starts with one question: which inference option did the request use? AWS now offers six, and the same model can cost half as much or 75% more depending on which you pick. The catch is that not every model supports every option.

Option	Price vs Standard	Commitment	Good for
Standard	baseline	None	Most synchronous traffic; the default, and the only option on every model
Priority	+75%	None	Customer-facing requests where latency matters most (Nova, gpt-oss and others; not Claude)
Flex	−50%	None	Work that can tolerate slower, variable response times (Nova, gpt-oss and others; not Claude)
Batch	−50%	None; async via S3, job timeout of 24 to 168 hours	Offline summarization, classification, evaluation
Reserved	fixed monthly	1 or 3 months	Predictable, mission-critical traffic; 99.5% uptime target
Provisioned Throughput	fixed hourly	None, 1 or 6 months	Custom models and workloads that need dedicated capacity

Let’s walk through each one. All examples below use Claude Sonnet 4.6 through a global cross-Region inference profile, called from US East (N. Virginia), at October 2026 prices, unless otherwise noted.

Standard (on-demand)

Standard is super straightforward: you pay for the input and output tokens your model chews through (or per image, or per second of video or audio). No commitment, no strings attached. This is where nearly everyone starts and where most production traffic stays.

Say you’re building a customer support chatbot. Here are our assumptions:

• Traffic: 10,000 conversations per day
• Input: 1,500 tokens per conversation (system prompt plus chat history), at $3.00 per 1M input tokens
• Output: 400 tokens per conversation, at $15.00 per 1M output tokens

And the cost breakdown:

• Input cost: 10,000 × 1,500 = 15M tokens/day × $3.00 = $45/day
• Output cost: 10,000 × 400 = 4M tokens/day × $15.00 = $60/day
• Total daily cost: $105/day
• Total monthly cost: $105 × 30 days = about $3,150/month

Notice that output is the bigger line even though there are far fewer output tokens. That holds for every model in the price table further down: output tokens cost roughly four to eight times more than input tokens, so the length of the answer, not the prompt, usually drives the bill.

Hold onto this $3,150 number. We’ll use it to show what each of the other options does to the same workload, and where an option doesn’t exist for Claude, we’ll say so.

Priority and Flex

AWS added these two on November 18, 2025, and they’re the easiest lever most teams haven’t pulled yet.

• Priority costs 75% more than Standard and puts your requests at the front of the queue. Use it for the one or two paths where a slow response costs you a customer.
• Flex costs 50% less than Standard and lets AWS process your request when capacity is available. Responses may be slower and more variable, which is fine for internal tools and background jobs.

Both share the same on-demand quota as Standard. One big caveat: as of October 2026, none of the Claude models on Bedrock support Priority or Flex. The model cards for Haiku 4.5, Sonnet 4.6, Sonnet 5, Sonnet 5.5, Opus 4.8, Opus 5.5, Haiku 5.5, and Fable 5.1 all list both tiers as unsupported. The same goes for OpenAI’s GPT-5.6 Terra. Amazon Nova Pro, Nova 2 Lite, and gpt-oss-120b do support them, so check the model card before you plan around either one.

So our Claude chatbot can’t go to Flex. Running the same traffic on Nova Pro shows what the tiers do (Nova Pro is $0.80 / $3.20 per 1M tokens on Standard):

• Standard: 15M × $0.80 = $12.00 plus 4M × $3.20 = $12.80, so $24.80/day, about $744/month
• Flex: 15M × $0.40 = $6.00 plus 4M × $1.60 = $6.40, so $12.40/day, about $372/month
• Priority: 15M × $1.40 = $21.00 plus 4M × $5.60 = $22.40, so $43.40/day, about $1,302/month

If some of that traffic is an internal helpdesk rather than customers, and Nova Pro passes your quality tests for it, moving just that slice to Flex halves its cost.

Batch

With batch inference, you submit a set of prompts as JSON Lines files in S3, and Bedrock writes the results back to S3 asynchronously. You set a job timeout between 24 and 168 hours, and you pay 50% of the Standard price on supported models.

A few caveats:

• Tool calling: It fails if you build the job in the Converse format (records with toolConfig return a 400 error). It works if you use the InvokeModel format with each model’s native request body.
• Model support: Batch isn’t supported on Claude Sonnet 5, Sonnet 5.5, Opus 4.8, Opus 5.5, Haiku 5.5, or Fable 5.1 as of October 2026. Among current Claude models, Haiku 4.5 and Sonnet 4.6 support it. Refer to the AWS docs for the current model list.

Suppose you want to summarize yesterday’s 10,000 support conversations overnight on Sonnet 4.6, each one 2,000 tokens in and 200 tokens out:

• Input cost: 20M tokens × $1.50 (batch rate) = $30
• Output cost: 2M tokens × $7.50 (batch rate) = $15
• Total: $45 per nightly batch, about $1,350/month

Doing that same job on Standard would cost 20M × $3.00 + 2M × $15.00 = $90 a night, twice as much. Batch is the cheapest way to run a frontier model on anything that doesn’t need an answer right now, as long as your model supports it.

Reserved Tier

Reserved launched in November 2025. You reserve input and output tokens-per-minute (TPM) capacity separately, for one or three months, at a fixed price per 1K TPM, billed monthly. Traffic above your reservation overflows to Standard automatically, and AWS targets 99.5% uptime for reserved traffic.

A few details worth knowing before you call your account team (Reserved is enabled through them, not the console):

• Minimums: 100,000 input TPM and 10,000 output TPM.
• Caching counts: If you use prompt caching, cache-write tokens count toward your TPM, so add CacheWriteInputTokens to InputTokenCount in CloudWatch when you size it.
• It doesn’t stop on its own: Billing continues until you delete the reservation with your AWS account team.
• Model support: It started with Claude Sonnet 4.5. Today the model cards list it for Sonnet 4.6 and Haiku 4.5, but not for Sonnet 5, Sonnet 5.5, Opus 4.8, or Opus 5.5.

Reserved makes sense once you have a few months of steady production traffic to size it against. Before that, you’re guessing, and you pay for the reservation whether you use it or not.

Provisioned Throughput

Provisioned Throughput is dedicated model capacity, measured in model units and billed hourly whether you use it or not. You can go no-commitment, 1-month, or 6-month; the longer the commitment, the better the rate.

To put numbers on “whether you use it or not”: one model unit of Nova 2 Lite in US East (N. Virginia) is $60.50/hour with no commitment, $55.00/hour on a 1-month term, and $30.25/hour on a 6-month term. Over a 730-hour month, that’s $44,165, $40,150, or $22,082.50 for a single model unit, busy or idle.

Custom and fine-tuned models used to require Provisioned Throughput, full stop. Since July 2025, Bedrock can also deploy a custom model for on-demand, pay-per-token inference. That currently covers fine-tunes of Nova Micro, Nova Lite, Nova Pro, and Nova 2 Lite in US East (N. Virginia) and Llama 3.3 70B Instruct in US West (Oregon), customized on or after July 16, 2025. Nova fine-tunes are billed at the base model’s token rate (for Nova 2 Lite, the $0.30 / $2.50 Global rate), and on the Nova side it’s for parameter-efficient (PEFT) fine-tunes only. Full-rank fine-tunes, older custom models, and anything that needs guaranteed throughput still need Provisioned Throughput.

For base models, the Provisioned Throughput list is short (mostly Amazon Nova, Titan, some Llama 3.x, and Claude 3 in a single Region), and AWS quotes most Provisioned Throughput pricing through your account team. For current Claude models, Reserved is the dedicated-capacity option, where the model supports it.

Prompt caching and cross-Region pricing

Two modifiers apply on top of whichever tier you pick.

Prompt caching stores a stable prefix, like your system prompt or a long reference document, so repeated requests pay a cache-read rate instead of the full input price. On most Claude models that’s 10% of the input price; Sonnet 5.5 and Opus 5.5 go down to 5%, and Fable 5.1 to 2.5%. Writing to the cache costs a premium (1.25× the input price for the 5-minute cache on Claude), so caching pays off once a prefix gets reused a few times within the cache window.

Back to our chatbot. Say 1,200 of those 1,500 input tokens are a system prompt that never changes (comfortably above Sonnet 4.6’s 1,024-token minimum per cache checkpoint). Here are our assumptions: cache reads at $0.30 per 1M tokens, cache writes at $3.75 per 1M, and, to be conservative, the cache gets rewritten once every 5 minutes, around the clock.

• Cached reads: 10,000 × 1,200 = 12M tokens × $0.30 = $3.60/day
• Uncached input: 10,000 × 300 = 3M tokens × $3.00 = $9.00/day
• Cache writes: 288 writes × 1,200 = 345,600 tokens × $3.75 = about $1.30/day
• Output: unchanged at $60/day
• Total: about $73.90/day × 30 = about $2,220/month, down from $3,150, with no change to the model or the answers

Cross-Region inference profiles come in two kinds. Geographic profiles (US, EU, and so on) route requests within a boundary and keep your data residency intact. Global profiles route anywhere. For Claude 4.5 and later, and other newer models such as Nova 2 Lite and GPT-5.6 Terra, Geo and in-Region requests cost 10% more than Global. Our chatbot through the US Geo profile would be 15M × $3.30 = $49.50 plus 4M × $16.50 = $66.00, so $115.50/day, about $3,465/month: $315 more for keeping the traffic in the US. Either way, the source Region sets the price.

Other charges to expect

A few more things can show up on a Bedrock bill:

• Guardrails, Knowledge Bases, Data Automation, Flows, and AgentCore: Each is metered separately. They’re out of scope for this guide, but if you’re running agents, expect them to be a meaningful share of the bill. One user request can trigger several model calls plus retrieval, memory, and tool charges.
• Model customization: Fine-tuning and distillation are billed per training token or training hour, plus $1.95 per model per month for storage. The resulting model runs on Provisioned Throughput or, for supported Nova and Llama 3.3 fine-tunes, on demand (see above).
• Model evaluation: You pay for the inference of the model under test plus judge-model tokens; human evaluation is $0.21 per completed task.

What do Bedrock models cost per million tokens?

The tier decides the multiplier; the model decides the base price. Here are the models most teams actually choose between, at Standard tier, from US East (N. Virginia), as of October 2026. Prices are the Global cross-Region rate where the model offers one; Geo and in-Region are 10% higher for the Claude, OpenAI GPT-5.6 and Nova 2 models. Nova Micro, Lite and Pro, gpt-oss-120b and Llama 4 Scout have a single rate. The last column shows which discounted or reserved options each model supports. Always check the AWS pricing page for your Region.

Model	Input per 1M tokens	Output per 1M tokens	Beyond Standard	Typical use
Amazon Nova Micro	$0.035	$0.14	Batch	Classification, routing, cheap text tasks
Amazon Nova Lite	$0.06	$0.24	Batch	Multimodal at low cost
Amazon Nova 2 Lite	$0.30	$2.50	Batch, Flex, Priority	Newer low-cost multimodal model
Amazon Nova Pro	$0.80	$3.20	Batch, Flex, Priority	Balanced general-purpose
Claude Haiku 5.5	$0.10*	$0.50*	Standard only	Fast, high-volume assistants
Claude Haiku 4.5	$1.00	$5.00	Batch, Reserved	Fast assistants on the previous generation
Claude Sonnet 4.6	$3.00	$15.00	Batch, Reserved	Production default for many teams
Claude Sonnet 5	$2.00	$10.00	Standard only	Balanced model, prior Sonnet
Claude Sonnet 5.5	$2.00	$10.00	Standard only	Newest balanced model
Claude Opus 5.5	$4.00	$20.00	Standard only	Deep reasoning and complex agents, at a lower price than Opus 4.8
Claude Opus 4.8	$5.00	$25.00	Standard only	Deep reasoning on the previous generation
Claude Fable 5.1	$10.00	$50.00	Standard only	Top-end model for long agent runs; needs the AWS-review data retention opt-in
OpenAI GPT-5.6 Terra	$2.00	$12.00	Standard only	Balanced OpenAI option (prompts up to 272K tokens)
OpenAI gpt-oss-120b	$0.15	$0.60	Batch, Flex, Priority	Open-weight general-purpose
Meta Llama 4 Scout	$0.17	$0.66	Batch	Open-weight, cost-sensitive workloads

*Claude Haiku 5.5 prices are for prompts up to 100,000 tokens. Longer prompts are billed at $0.50 / $2.50.

One more note on Fable 5.1: it’s only offered through the US Geo and Global profiles, and you have to opt your account in to AWS review data retention (aws_review) before you can call it. Its model card also warns of materially higher refusal rates. Requests blocked before inference starts aren’t billed, but a response cut off partway is billed for the tokens it generated before the block.

Three things jump out. First, the gap between the cheapest and the most expensive model is over 100×: $0.14 versus $50 per million output tokens is more than 350×. Second, our $3,150/month chatbot would cost about $744/month on Nova Pro (we did the math in the Flex section), or 15M × $0.30 + 4M × $2.50 = $14.50/day, about $435/month on Nova 2 Lite, if either passes your quality tests. Picking the right model for each task matters more than any discount AWS offers.

Third, per-token prices across Claude generations aren’t quite apples to apples. Claude 4.7 and later models use a newer tokenizer that, per Anthropic, produces roughly 30% more tokens for the same text (the exact increase depends on your content). Sonnet 5 looks a third cheaper than Sonnet 4.6, but our chatbot’s 15M input and 4M output tokens become about 19.5M and 5.2M. That’s 19.5M × $2.00 = $39 plus 5.2M × $10.00 = $52, so $91/day, about $2,730/month. That’s about 13% less than $3,150, not the $2,100 you’d get by swapping prices alone. Measure your own token counts before you budget the switch.

How to estimate and track what Bedrock actually costs you

The examples above are the template for estimating. Here’s how to do it for real, and then how to keep an eye on the number once traffic is flowing.

AWS Pricing Calculator

Start with the AWS Pricing Calculator for a rough estimate before you even have an account. You enter expected input and output tokens per month per model and it produces a number. Its Bedrock coverage has historically lagged for third-party models, so for Anthropic or OpenAI models, grab the per-million rates from the pricing table above and run the math yourself. It’s a few multiplications.

AWS Cost Explorer

Once you’re running, Cost Explorer is the fastest way to see where the money goes, with one gotcha that catches a lot of people.

Amazon’s own models, like Nova, show up under the Amazon Bedrock service. Third-party models such as Claude are offered and billed through AWS Marketplace, so their charges appear under the model’s own product name, for example Claude Sonnet 4.6 (Amazon Bedrock Edition), and not under Amazon Bedrock. Filter Service to “Amazon Bedrock” alone and your Claude bill will look suspiciously small.

So:

1. Open Cost Explorer and set Group by to Service.
2. Look for Amazon Bedrock and for each model’s Marketplace product name, then filter to the ones you use.
3. Switch Group by to Usage type. The usage type encodes the Region, the token type, and the routing, for example USE1-NovaPro-input-tokens for Nova Pro, or USE1-MP:USE1_InputTokenCount_Global-Units for global input tokens on Claude Sonnet 4.6. Cache reads, cache writes, and batch get their own usage types.

Use monthly granularity for the trend and daily when something spikes. Cost Explorer also forecasts up to 18 months ahead, which gets useful once you have a few months of history. If you’re weighing Cost Explorer against Budgets for this, we compared them in AWS Budgets vs. Cost Explorer.

Tags, inference profiles, and Budgets

Bedrock invocations can be attributed to an application inference profile and, since April 2026, to the IAM principal that made the call (extended to the bedrock-mantle endpoint in August 2026). That’s what makes team-level chargeback possible. Tag your inference profiles and IAM roles by team and environment, activate those tags as cost allocation tags, then set an AWS Budget with an alert at 80% of what you expect.

One more: model invocation logging is off by default. Turn it on if you want per-call token counts to reconcile against the bill.

Cost Optimization Strategies for AWS Bedrock

Before we dive into cost-saving tips, let’s quickly highlight a few challenges that can make AWS Bedrock pricing tricky to manage:

• Dynamic usage patterns: Your workloads can fluctuate, causing unpredictable costs.
• Complexity of cost structures: There are several pricing models and different rates for different foundation models (FMs), and not every model supports every tier.
• Visibility into costs: Without proper monitoring, it’s tough to see exactly where your money is going, especially when half of it is billed through Marketplace.

The good news is you can tackle all of these issues with a few simple strategies.

Monitoring Your Workloads

Want to optimize your Bedrock costs? Start by keeping an eye on your usage. Amazon CloudWatch publishes Bedrock metrics such as InputTokenCount and CacheWriteInputTokens, plus a ResolvedServiceTier dimension that shows which tier actually served each request. Here are some essential features of CloudWatch to take advantage of:

• Custom dashboards: Track the key metrics that matter the most to your application. This can be input/output tokens, or model performance.
• Alarms: Set a few usage thresholds so that CloudWatch notifies you before things get too expensive.

Additionally, you can use AWS CloudTrail to log API calls. This gives you a handy audit trail that shows who is using what resources, and how. Stay on top of your workloads, and those surprise charges won’t have a chance to sneak up on you.

Optimizing Model Usage

Bedrock offers a wide selection of FMs. While it’s tempting to go for the “best” or the “cheapest”, the real key is finding the most cost-effective option for your specific use case.

Not all models are created equal. Some are total overkill for certain applications. For example, if you’re working on a simple text classification task, there’s no need to use a complex, expensive model. Instead, choose a less complex FM that’s cheaper and better suited to your needs. This saves you both processing and token costs, without sacrificing too much on the quality your application requires.

By matching your model choice to the complexity of your task, you can get the job done efficiently and avoid unnecessary expenses.

Using Flex, Reserved, and Provisioned Throughput

If you’re on a model that supports it, the biggest easy win is moving anything that isn’t customer-facing to the Flex tier for a 50% discount. Internal tools, background enrichment, and evaluation runs rarely need Standard-tier latency. On Claude, Flex isn’t available, so your levers are prompt caching, Batch (on Haiku 4.5 and Sonnet 4.6), and a cheaper model for the background work.

For larger, steady workloads, the Reserved tier gives you dedicated tokens-per-minute capacity at a fixed monthly price, with overflow to Standard, on the models that support it. Provisioned Throughput still exists, but in 2026 it’s mostly for custom models that can’t use on-demand deployment or need guaranteed throughput.

Either way, the only way to know for sure whether a reservation is worth it is by monitoring your usage for a few months first. That gives you the data to size it right instead of guessing.

Caching Your Prompts

If your requests share a long, stable prefix, like a system prompt, a tool definition, or a reference document, turn on prompt caching. Cached reads cost around 10% of the normal input price on most Claude models, 5% on Sonnet 5.5 and Opus 5.5, and 2.5% on Fable 5.1. Check the model card for the minimum prefix length per cache checkpoint: it ranges from 512 tokens (Opus 5.5, Fable 5.1) to 4,096 (Haiku 4.5). Put the stable content first and the user’s message last so the cache actually hits. Track cache reads versus writes in Cost Explorer; if writes dominate, your prefix isn’t stable enough to cache.

Batch Processing for Cost Efficiency

For larger datasets, batch processing is a simple way to cut costs. Rather than sending requests one at a time, you submit the whole dataset as JSON Lines files in S3 and pay half the Standard rate on supported models.

This works best for jobs like sentiment analysis or translation, where nothing needs an answer right away. You use the same number of tokens; you just pay less for each one. Check that your model supports Batch first, since several of the newest Claude models don’t.

AWS Tags and Cost Categories for Cost Visibility

To get a clearer picture of AWS Bedrock expenses, start by tagging your inference profiles and IAM roles, then organize them with AWS Cost Categories. Tagging lets you track costs by project, team, or environment, so each team can see its own share.

With AWS Cost Categories, you can then group those tagged costs into spending buckets that match how your business is organized. That’s also a tidy way to pull Amazon Bedrock and the Marketplace model charges into a single “Generative AI” bucket.

We wrote up two helpful guides on tagging resources and also cost categories:

1. AWS Cost Categories: A Better Way to Organize Costs
2. AWS Tags Best Practices and AWS Tagging Strategies

Conclusion

AWS Bedrock is a powerhouse for building and scaling generative AI applications, but the pricing has more moving parts every year. The good news is that the big levers are simple: pick the cheapest model that passes your quality bar, cache your prompts, move anything that isn’t customer-facing to Flex or Batch where your model supports them, and only reserve capacity once you’ve measured a few months of real traffic. Moving our chatbot from Sonnet 4.6 to Nova Pro alone takes it from $3,150 to about $744 a month, roughly a quarter of the day-one cost.

Now jump into Cost Explorer, group by Service, and find out where your Claude charges have been hiding.

If you’d rather not build those views yourself, also consider CloudForecast. It sends a daily AWS cost report to Slack, Microsoft Teams, or email with what changed yesterday and why, down to service, tag, and resource, plus anomaly alerts when spend jumps. That includes Claude and other third-party models billed through AWS Marketplace, so they sit right next to the rest of your Bedrock spend instead of hiding under their own product names. You can try it free for 30 days.

Frequently asked questions about AWS Bedrock pricing

How much does AWS Bedrock cost?

There’s no flat fee. You pay per token (or per image or second of media) for the model you use. As of October 2026, prices range from $0.035 per million input tokens for Amazon Nova Micro to $50 per million output tokens for Claude Fable 5.1. A support chatbot handling 10,000 conversations a day on Claude Sonnet 4.6 costs roughly $3,150 a month at Standard rates, about $2,220 with prompt caching, and about $744 if you can move it to Amazon Nova Pro. Claude models don’t support the Flex tier.

Does Amazon Bedrock have a free tier?

Not a model-specific one, but new AWS accounts created since July 2025 get $100 in credits at sign-up and can earn up to $100 more by trying services including Bedrock. On the free account plan, the credits run out after six months or when they’re used up, whichever comes first, and they apply to Bedrock usage. There’s no ongoing free monthly allowance of tokens.

How much does Claude cost on AWS Bedrock?

As of October 2026, at Standard tier through a global cross-Region profile: Claude Haiku 5.5 is $0.10 input / $0.50 output per million tokens (for prompts up to 100,000 tokens), Haiku 4.5 is $1 / $5, Sonnet 4.6 is $3 / $15, Sonnet 5 and Sonnet 5.5 are both $2 / $10, Opus 5.5 is $4 / $20, Opus 4.8 is $5 / $25, and Fable 5.1 is $10 / $50. Geo and in-Region profiles cost 10% more. Batch is half price on Haiku 4.5 and Sonnet 4.6; none of these Claude models support Flex or Priority. Claude 4.7 and later models also produce about 30% more tokens for the same text, so compare costs on your own prompts.

Is there an AWS Bedrock pricing calculator?

Yes. The AWS Pricing Calculator has a Bedrock section where you enter monthly input and output tokens per model. Coverage of third-party models is uneven, so for those, multiply your expected tokens by the per-million rates on the Bedrock pricing page. The examples in this guide show the math.

What is Bedrock provisioned throughput?

Provisioned Throughput is dedicated model capacity, sold in model units and billed by the hour with no-commitment, 1-month, or 6-month terms. You pay for the capacity whether you use it or not. It’s required for full-rank fine-tunes and other custom models that can’t use on-demand deployment (supported Nova and Llama 3.3 fine-tunes can now run on demand), and it’s only available for a short list of base models; for current Claude models, the Reserved tier is the dedicated-capacity option where it’s supported.

Is Bedrock more expensive than calling Anthropic or OpenAI directly?

For most models, Bedrock’s global Standard per-token price matches the provider’s own API price. What you get for it is running inside your AWS account: IAM, VPC endpoints, CloudTrail, and consolidated billing.

Third-party models such as Claude and GPT are billed through AWS Marketplace. AWS Marketplace’s Claude listing says that spend draws down against an existing EDP or PPA commitment. But it’s billed at the provider’s price (your EDP discount doesn’t apply to it), and Marketplace drawdown is subject to your agreement’s eligibility rules and cap. Get it confirmed in writing by your account team before you count on it. We cover the Marketplace rules in our AWS EDP guide.

Geo and in-Region endpoints for Claude 4.5 and later, and for many newer models, carry a 10% premium over global pricing.

What are Bedrock’s token limits?

Limits are quotas, not prices. Each model and Region has a tokens-per-minute quota on the on-demand tiers (and, for many models, a requests-per-minute quota), visible in Service Quotas, and you can request increases. Reserved and Provisioned Throughput give you capacity outside those quotas. Some models also apply a burndown multiplier to quota consumption (GPT-5.6 Terra counts each output token as 10 on bedrock-runtime), which is separate from what you’re billed.