Cost levers - verified
Prompt caching: automatic and explicit. Cache lifetime: 5 min default; 1 hour on Claude 4.5+; 30 min on OpenAI GPT-5.6/6.1. A cache hit typically saves 75% on input across 4 current Amazon models. Batch: 50% off select models, 24 to 168 hour job timeout. Batch and caching stack: not stated. Service tiers: Flex (50% off), Priority (75% premium), Reserved (fixed price per 1K TPM, monthly).
Read from docs.aws.amazon.com, aws.amazon.com. Compare every provider's cost levers
Both: automatic caching by default, plus explicit breakpoints you control
Implicit Prompt Caching (best effort, no controls) and Explicit Prompt Caching (cache checkpoints: cachePoint in Converse, cache_control for Claude InvokeModel, prompt_cache_breakpoint for OpenAI models). Support varies by model and API. Simplified cache management looks back about 20 content blocks from a single checkpoint.
Resets on each hit. Claude 4.5+ and Haiku 4.5: 5 minutes or 1 hour; Claude 3.x and Amazon Nova: 5 minutes; OpenAI GPT-5.6/6.1 models: 30 minutes minimum. Default 5 minutes when no ttl is given.
Per checkpoint, cumulative prefix: ranges from 512 to 4,096 by model (Claude 5.x family 512; Opus 4.8 / Sonnet 5 / Sonnet 4.x / Claude 3.x 1,024; Opus 4.5-4.7 and Haiku 4.5 4,096; Nova 1K; OpenAI GPT-5.6/6.1 1,024)
Model-dependent: writes can be billed above standard input; OpenAI models on Bedrock bill writes at 1.25x uncached input
Model's cache-read rate
Add a cachePoint (Converse) or cache_control (Claude InvokeModel) checkpoint after static content; up to 4 checkpoints for Claude; add "ttl": "1h" where supported.
"cachePoint": {"type": "default", "ttl": "1h"}“Amazon Bedrock supports two types of prompt caching: Implicit Prompt Caching and Explicit Prompt Caching. Support for each type varies by model and API.”
“Implicit Prompt Caching is best effort. Repeating an identical prompt doesn't guarantee a cache hit, and cache-hit rates can vary.”
“The cache has a Time To Live (TTL), which resets with each successful cache hit.”
“Many models support a 5-minute TTL. Check the model card for your model to see the exact TTL conditions.”
“If no `ttl` value is provided, the default 5-minute caching behavior applies.”
“Cached prefixes remain available for reuse for at least 30 minutes, long enough to cover the burst of calls a single agent run generates.”
“Minimums currently range from 512 to 4,096 tokens per cache checkpoint.”
“Depending on the model, tokens written to cache can be billed at a rate that is higher than the standard input token rate.”
“Tokens written to cache are billed at 1.25× the uncached input token rate.”
“"cachePoint" : { "type": "default", "ttl" : "1h" }”
“When you enable simplified cache management, the system automatically checks for cache hits at previous content block boundaries, looking back up to approximately 20”
Sources: docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html, aws.amazon.com/bedrock/pricing
50% off select models
50% lower price than on-demand (select foundation models)
No fixed completion window stated; the job timeout is set per job with timeoutDurationInHours, 24 to 168 hours (API reference). No default is stated on the cited pages.
Per-model quotas in Amazon Bedrock endpoints and quotas (AWS General Reference, not stated on the batch pages); job timeout 24 to 168 hours via timeoutDurationInHours; not supported for provisioned models
Batch inference job (model invocation job) with S3 input/output; OpenAI Batch API also supported
Not stated: the provider's pages do not say
“Amazon Bedrock offers select foundation models (FMs) from leading AI providers like Anthropic, Meta, Mistral AI, and Amazon for batch inference at a 50% lower price compared to on-demand inference pricing.”
“To see quotas for batch inference, see Amazon Bedrock endpoints and quotas in the AWS General Reference.”
“timeoutDurationInHours | To specify the duration in hours after which the job will time out.”
“Batch inference isn't supported for provisioned models.”
“The number of hours after which to force the batch inference job to time out.”
“Valid Range: Minimum value of 24. Maximum value of 168.”
Sources: docs.aws.amazon.com/bedrock/latest/userguide/batch-inference.html, docs.aws.amazon.com/bedrock/latest/userguide/batch-inference-create.html, docs.aws.amazon.com/bedrock/latest/APIReference/API_CreateModelInvocationJob.html, aws.amazon.com/bedrock/pricing
Flex trades speed or availability for a lower price; Priority and Fast cost more for lower latency or reserved capacity.
Flex
Set with service_tier: flex
50% discount to Standard
Longer processing times; shares on-demand quota
“Flex tier pricing is at 50% discount to Standard tier pricing”
“For workloads that can handle longer processing times, the Flex tier offers cost-effective processing for a pricing discount.”
Sources: aws.amazon.com/bedrock/pricing, docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
Priority
Set with service_tier: priority
75% premium to Standard
Fastest response times; prioritized over Standard and Flex; no reservation
“Priority tier pricing is at 75% premium to Standard tier pricing”
“The Priority tier delivers the fastest response times for a price premium over standard on-demand pricing.”
Sources: aws.amazon.com/bedrock/pricing, docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
Reserved
Set with service_tier: reserved
Fixed price per 1K tokens-per-minute, billed monthly; 1 or 3 month terms
Reserved prioritized capacity, 99.5% uptime target, overflow to Standard; minimum 100,000 input TPM and 10,000 output TPM
“Customers can reserve capacity for 1 month or 3 month duration. Customers pay a fixed price per 1K tokens-per-minute and are billed monthly.”
“The Reserved tier targets 99.5% uptime for model response.”
“The Reserved tier provides the ability to reserve prioritized compute capacity for your mission-critical applications that cannot tolerate any downtime.”
“Batch inference isn't supported for provisioned models.”
USD per 1M tokens from the provider's own pricing page, newest release first. A dash means the model has no such rate. These are the Amazon models billed through Amazon Bedrock. Muted rows are scheduled for shutdown or deprecated.
| Model | Input | Cached input | Batch input | Output | Batch output | Fast modeinput / output |
|---|---|---|---|---|---|---|
| Amazon Nova 2 Lite | $0.30 | $0.07575% off | $0.15 | $2.50 | $1.25 | - |
| Amazon Nova Premier | $2.50 | - | $1.25 | $12.50 | $6.25 | - |
| Amazon Nova Lite | $0.06 | $0.01575% off | $0.03 | $0.24 | $0.12 | - |
| Amazon Nova Micro | $0.035 | $0.0087575% off | $0.0175 | $0.14 | $0.07 | - |
| Amazon Nova Pro | $0.80 | $0.2075% off | $0.40 | $3.20 | $1.60 | $1 / $4 |
Batch completion window is not stated on the Bedrock batch overview page; the per-job timeout range (24 to 168 hours) comes from the CreateModelInvocationJob API reference.
Both: automatic caching by default, plus explicit breakpoints you control. Implicit Prompt Caching (best effort, no controls) and Explicit Prompt Caching (cache checkpoints: cachePoint in Converse, cache_control for Claude InvokeModel, prompt_cache_breakpoint for OpenAI models). Support varies by model and API. Simplified cache management looks back about 20 content blocks from a single checkpoint. How to enable: Add a cachePoint (Converse) or cache_control (Claude InvokeModel) checkpoint after static content; up to 4 checkpoints for Claude; add "ttl": "1h" where supported.
Across 4 current models with a cache read price, a hit saves 75% off the input price. The docs say: Model's cache-read rate.
50% lower price than on-demand (select foundation models). Window: No fixed completion window stated; the job timeout is set per job with timeoutDurationInHours, 24 to 168 hours (API reference). No default is stated on the cited pages.. Stacks with caching: not stated: the provider's pages do not say.
Flex: 50% discount to Standard. Priority: 75% premium to Standard. Reserved: Fixed price per 1K tokens-per-minute, billed monthly; 1 or 3 month terms.
Last verified from Amazon Bedrock's own pages below. Quotes are verbatim; each was matched against the live page.
Newsletter
One email per week. New model releases, pricing changes and what they mean for what you build.