Amazon Bedrock does not have one price list. It has many meters, and a single request can touch several: a model, a cache, a retrieval index, a safety check, and a tool. Each meter has its own unit, the unit can differ between models within a modality, and rates depend on the model, provider, and Region. This article explains how Bedrock pricing is organized and how charges are calculated. It quotes no prices, because rates vary and change.
How Amazon Bedrock Pricing Is Structured
Bedrock pricing is easier to read when you separate five layers that people often treat as one:
- Billing meters. The unit being counted: tokens, images, queries, text units, model units, or hours.
- Service tiers. Standard, Priority, and Flex operate within the on-demand framework and set how requests are prioritized and priced. Reserved is also a service tier, but it works as a capacity reservation with its own monthly billing.
- Inference methods. On-demand for synchronous requests, and batch for asynchronous jobs.
- Routing configuration. Cross-region inference profiles, which decide where a request can be processed.
- Dedicated capacity. Provisioned Throughput, a purchase of model units for one model, billed hourly.
These layers are not interchangeable plans, and each is documented separately. Pricing, tier availability, inference profiles, caching, customization options, and capacity commitments can differ by model and Region. Feature meters, such as customization, Knowledge Bases, and Guardrails, bill on their own units beside this stack.
Pricing Categories at a Glance
| Category | What it is | Primary billing unit | Commitment |
| Standard | On-demand service tier, default | Per input and output token, or the unit for the modality | None |
| Priority | On-demand service tier | Token rates at a premium over Standard | None |
| Flex | On-demand service tier | Token rates at a discount to Standard | None |
| Batch | Asynchronous inference method | Batch rates for supported models | None |
| Cross-region inference | Routing configuration | Rate for the profile type, based on the source Region | None |
| Reserved | Service tier and capacity reservation | Fixed price per 1K tokens per minute, billed monthly | 1 or 3 months |
| Provisioned Throughput | Dedicated capacity purchase | By Model Units: per unit per hour. A token-based type also exists | None, 1 month, or 6 months |
| Customization | Model lifecycle | Training, storage, and hosting or inference, by method | None |
| Knowledge Bases | Feature meter | Managed: storage and retrieval. Customer-managed parts bill separately | None |
| Guardrails | Feature meter | Text units for most policies. Automated Reasoning has its own meter | None |
Model Inference: On-Demand Pricing
On-demand inference serves synchronous requests with no commitment. The billing unit depends on the model and modality, and not every model supports every modality:
- Text models bill per input token and per output token.
- Embedding models bill per input token, although some listings, such as certain Cohere Embed models, are offered through Provisioned Throughput.
- Image models bill per image or per generation, and some vary by resolution, quality, step count, or other model-specific dimensions.
- Reranking bills per query, where one query holds up to 100 document chunks and larger requests count as multiple queries.
- Speech, video, and multimodal models have model-specific pricing entries, and some speech workloads add text token charges, such as for transcription.
For text models, a request's charge is input tokens times the input rate plus output tokens times the output rate, specific to a model, Region, and tier. Reasoning tokens generally bill as output tokens.
You select a tier per request with the service_tier parameter, and omitting it routes the request to Standard.
- Standard is the default and baseline, available across foundation models.
- Priority is priced at a premium over Standard and receives preferential queue treatment.
- Flex is priced at a discount to Standard, and AWS notes it is processed after Standard requests during high traffic.
Priority and Flex are not offered for every model, and some are limited to specific Regions. On-demand quota for a model is shared across these three tiers.
Batch Inference Pricing
Batch inference is an asynchronous method. You place input files in Amazon S3, submit a job, and retrieve output files from S3. Inputs use the InvokeModel or Converse format. Each record is handled as a separate request, output order is not guaranteed to match input order, and any conversation history must be included in the record. For selected foundation models, AWS lists batch inference at a 50% lower price than on-demand inference. Availability and pricing depend on the model and Region.
AWS documents these limitations:
- Batch inference is not supported for provisioned models.
- Converse-type jobs do not support tool calling, and records with toolConfig fail with a 400 error. AWS points to InvokeModel-type jobs, using the model's native format, for tool calling.
- Structured outputs are documented as supported with both invocation types, but confirm support for your model.
- Prompt caching is not available in batch.
- On-demand requests take priority over batch jobs, and jobs have a configurable timeout.
Beyond these, support is model-specific, so check the batch model and Region list.
Cross-Region Inference Pricing
Cross-region inference routes requests to other AWS Regions through an inference profile. It is a routing configuration and does not replace a tier or a commitment. Geographic profiles route requests within a configured geography, such as the US, EU, APAC, Japan, or Australia, depending on the model and source Region. Global profiles can route to any supported commercial Region.
AWS states there is no additional routing charge. Pricing is calculated from the source Region, the Region you call from, while processing can occur in any destination Region in the profile. Geographic profiles permit processing in other Regions within the geography, and destination Regions can vary by source Region, so verify them with GetInferenceProfile if data residency matters. Rates differ between in-region and cross-region variants. AWS documents global pricing only relative to geographic profiles.
Inference profiles do not support Provisioned Throughput. For some models, output tokens count at a higher multiple against quotas, which affects quota use, not billed tokens.
Reserved Tier Pricing
Reserved is both a service tier and a capacity reservation. You select it per request with the service_tier value reserved, but unlike Priority, Standard, and Flex it is not billed per token. You reserve input and output tokens per minute and pay a fixed price per 1K tokens per minute, billed monthly. Terms are one or three months, arranged through your AWS account team.
Reserved capacity is separate from on-demand quota. AWS documents minimum allocations of 100,000 input and 10,000 output tokens per minute, and a 99.5% uptime target for model responses, a target, not a stated guarantee. Requests beyond the reservation overflow to the Standard tier. AWS does not detail overflow billing, so verify the applicable Standard rate for your model and Region. Cache write input tokens count against a reservation, and billing continues until your account manager deletes it.
Provisioned Throughput Pricing
AWS documents two Provisioned Throughput types, by Model Units and by Tokens. They have their own purchase terms, and model availability and commitment rules depend on the option. This section covers model units. Check the token-based documentation before assuming the same mechanics.
By Model Units, you buy dedicated capacity for one specific model. A model unit defines how many input and output tokens the model can process per minute, and its size varies by model. Billing is hourly and depends on the model, the number of model units, and the commitment term: none, one month, or six months, with longer terms carrying a discounted hourly rate. A commitment cannot be deleted early, and hourly billing continues until deletion. It applies to base and custom models. New accounts start with no model units for committed purchases, so you request them through AWS Support. No-commitment support varies by base model.
Reserved versus Provisioned Throughput. Reserved is a per-request tier measured in tokens per minute. Provisioned Throughput by Model Units is a dedicated hourly purchase for one model.
Prompt Caching and Token Billing
Prompt caching reuses a repeated prompt prefix, such as a system prompt or reference document, and changes how input tokens are classified:
- Cache writes bill at the model's cache-write rate, which can be higher than the standard input rate.
- Cache reads bill at the cache-read rate.
- Other input tokens bill at the standard input rate.
With the Converse API, inputTokens excludes cached tokens, so total input equals inputTokens plus cacheReadInputTokens plus cacheWriteInputTokens.
Mechanics vary by model and provider. A model can support implicit caching, explicit caching, or both, and explicit caching needs cache checkpoints. Lifetime commonly defaults to five minutes and resets on each hit, but supported durations differ, including one-hour and thirty-minute windows. Minimum prefix sizes range from 512 to 4,096 tokens per checkpoint, and a shorter prefix is not cached. Cache entries are tied to the processing server, so an identical prompt routed elsewhere may cause a new write, not a read. Hits are not guaranteed, so check the response usage fields, and verify TTL behavior in the model's documentation. Caching applies to on-demand inference and not to batch.
Model Customization Pricing
Customization can add training, storage, hosting, and inference charges. Which apply depends on the customization method, the model, and the deployment option the model supports.
- Fine-tuning and continued pre-training. Training bills on volume: tokens trained (corpus tokens multiplied by epochs) for several text models, and images seen for some image models. Storage bills monthly per custom model. Inference runs on Provisioned Throughput, billed hourly per model unit, or through on-demand custom model deployment, billed for usage, where supported. On-demand deployment covers supported models, such as Amazon Nova models customized on or after July 16, 2025, so confirm model and Region eligibility. Some models are offered only through Provisioned Throughput.
- Reinforcement fine-tuning. Training bills per training hour for the models the pricing page lists. Inference on the resulting model bills per token on-demand.
- Custom Model Import. There is no import charge. Billing is per Custom Model Unit per minute, in five-minute windows, from the first successful invocation, plus monthly storage per unit. Units depend on architecture, parameter count, and context length, and copies scale to zero after five minutes without requests, so cold starts can occur.
Other Bedrock Capabilities
Knowledge Bases. AWS distinguishes Bedrock Managed Knowledge Bases from customer-managed ones. A managed Knowledge Base bills raw-data index storage per GB each month and standard retrieval per 1,000 calls, and managed parsing, embeddings, and reranking are included. Custom embedding, reranking, or planning models add their own model charges. Agentic retrieval bills for the agentic call plus each underlying retrieval, and AgentCore Gateway and CloudWatch observability can add separate charges. In customer-managed setups, your vector database, embedding and reranking models, and supporting infrastructure bill separately. Self-managed reranking bills per query, structured data retrieval per SQL generation request, and generated answers as inference.
Guardrails. Guardrails does not use one meter. Most policies bill per 1,000 text units, where a unit covers up to 1,000 characters and longer inputs split into multiple units. Contextual grounding counts the combined characters of source, query, and response. Image filters bill per image, and word and regex-based sensitive information filters are free. Automated Reasoning checks use a billing mechanism separate from standard policies. AWS documents charges for processed validation requests regardless of result, and the pricing page expresses the charge per 1,000 text units per Automated Reasoning policy. Refer to the current pricing schedule for the calculation. The InvokeGuardrailChecks API has separate rates.
Built-in tools and related features. Web search bills per call, AgentCore Gateway tool invocations carry their own charges, Intelligent Prompt Routing bills per request on top of inference, and Data Automation depends on output type. Model evaluation bills human review per completed task and judge models at on-demand token rates. Agents, Flows, and Marketplace models have their own pricing sections.
Additional Charges and Billing Visibility
Some charges fall outside the Bedrock meters. Runtime APIs are stateless, so multi-turn conversations resend history as input on each request, and agent workflows can involve several model and tool calls per user request. S3, CloudWatch, and any vector store or compute you run bill separately. Invocation logging is off by default, and logs and operational metrics are not billing records, so use the Cost and Usage Report for billed amounts.
Key Takeaways
- Service tiers, inference methods, routing configurations, and dedicated capacity are distinct layers and not interchangeable plans.
- Reserved is both a tier and a monthly capacity reservation. Provisioned Throughput is a separate dedicated purchase, by Model Units or by Tokens.
- Text inference bills input and output tokens at model-, Region-, and tier-specific rates, and caching splits input into write, read, and standard tokens.
- Customization billing depends on method, model, and deployment option, and Guardrails policies do not share one meter.
Frequently Asked Questions (FAQs)
Q1. How is Amazon Bedrock priced?
Each capability bills on its own meter, such as tokens, images, queries, text units, model units, or hours, at model- and Region-specific rates.
Q2. Are Standard, Priority, Flex, and Reserved the same kind of option as Batch or Provisioned Throughput?
No. Standard, Priority, and Flex are on-demand tiers. Reserved is a tier that works as a monthly capacity reservation. Batch is an asynchronous method, cross-region inference is routing, and Provisioned Throughput is a dedicated purchase.
Q3. Does cross-region inference have its own charge?
AWS states there is no additional routing charge. The price is calculated from the source Region, and the pricing page lists in-region and cross-region variants.
Q4. How are cached tokens billed?
Cache writes, cache reads, and other input tokens bill at separate rates, and cache behavior varies by model.
Q5. How are custom models billed?
Training bills on volume or hours and storage bills monthly. Inference depends on the method, model, and deployment option: Provisioned Throughput, on-demand deployment where supported, per-token billing for reinforcement fine-tuned models, or Custom Model Units for imports.

