Choosing a Bedrock Pricing Mode

Look forPay per model token with no commitment. Standard prices vary by provider and model; use the representative table above, then verify the exact model row and routing profile in AWS's pricing page.

Look forUse batch for S3-based jobs when the model lists a batch price. AWS advertises 50% below on-demand for supported models, and the supported model table changes by provider and release.

Look forChoose Flex for work that tolerates longer processing, at half the Standard price. Priority costs 75% more for customer-facing work that needs the fastest responses. Both share on-demand quotas. Check support on the model card.

Look forSize Reserved capacity on input plus cache-write tokens. It starts at 100,000 input and 10,000 output tokens per minute for 1 or 3 months. Overflow goes to Standard, with 99.5% target uptime. Ask the AWS account team about supported models, including Sonnet 4.6, Opus 4.6, and Haiku 4.5.

Look forBuy hourly Model Units with no commitment or with 1-month or 6-month terms. Custom models require Provisioned Throughput. AWS supplies Model Unit prices per supported model. Inference profiles do not support it.

Look forMeasure cache reads against uncached input. AWS prices cache reads below uncached input on supported model rows, and cache writes have their own rate. Put stable prompt prefixes behind explicit cache points and track reads per workload.