Four Amazon Bedrock inference choices that affect cost
Machine Learning Blog
This article explains four Amazon Bedrock inference request choices that impact token costs and processing without changing the model or prompt.
- Prompt cache duration: 5-minute caches pay back on first reuse; 1-hour caches are cheaper from second request within the hour if gaps exceed 5 minutes
- Service tier: Flex tier offers lower rates for latency-tolerant workloads like batch jobs; support varies by model and region
- Reasoning effort: Lowering from high to low reduces output tokens by ~169% on agent tasks; accuracy depends on task type and model
- maxTokens: Set per-task ceiling based on P95/P99 output distribution to limit spend and prevent truncation cascades in multi-turn chains
- Partition tenant-specific prefixes with SHA-256 hashes to prevent cache sharing across tenants in multi-tenant applications
- Validate model support for each setting before deployment; track cache reads, service tier resolution, and stop reasons in monitoring
Systematically test each setting one at a time, measure token usage and quality impact, and maintain capability records for each model version and region.
The AWS News Feed is currently looking for gold sponsors. If you want to support the AWS community and reach a large audience of AWS professionals, consider sponsoring the AWS News Feed.
Related articles
Jun 10
2025
2025
Effective cost optimization strategies for Amazon Bedrock
Oct 30
2026
2026
Managing Amazon Bedrock costs end to end: A playbook for FinOps and platform teams
Nov 12
2026
2026
Optimizing cost per task and proving agent ROI on Amazon Bedrock
Oct 27
2026
2026
Fix evaluation bias to unlock better models on Amazon Bedrock
The AWS News Feed is currently looking for silver sponsors. If you want to support the AWS community and reach a large audience of AWS professionals, consider sponsoring the AWS News Feed.