GUIDE · MONEY · #3515
Amazon Bedrock: four inference-level settings that affect cost
AWS details how four request-level choices—cache duration, service tier, reasoning effort, and maxTokens—change how many tokens Bedrock processes and which rates apply, so inference costs can shift even after prompt tuning. The post shows how to test and verify effects via response metadata, logs, and billing, cites TTL guidance for supported models (e.g., 5‑minute and 1‑hour TTLs; 30‑minute minimum for some GPT‑5.6 builds), and gives cache read/write pricing examples for Anthropic models.
KEY POINTS
- AWS details how four request-level choices—cache duration, service tier, reasoning effort, and maxTokens—change how many tokens Bedrock processes and which rates apply, so inference costs can shift even after prompt tuning.
- The post shows how to test and verify effects via response metadata, logs, and billing, cites TTL guidance for supported models (e.g., 5‑minute and 1‑hour TTLs; 30‑minute minimum for some GPT‑5.6 builds), and gives cache read/write pricing examples for Anthropic models.
- These request-level settings can materially change Bedrock bills independent of model choice or prompt quality, so operators need to monitor TTLs, tier selection, and token accounting to control costs.
WHY IT MATTERS
These request-level settings can materially change Bedrock bills independent of model choice or prompt quality, so operators need to monitor TTLs, tier selection, and token accounting to control costs.