What changed
Amazon Bedrock allows users to adjust inference settings that influence cost without changing the model or prompt. These include cache duration, service tier, reasoning effort, and maxTokens. Each choice affects how many tokens are processed and the rate applied, depending on the model and workload.
What this means in practice
The cost of inference on Amazon Bedrock can increase based on specific request-level choices. For example, cache duration settings for supported models like Anthropic Claude Sonnet 5 and Opus 5 affect how often tokens are billed.
A 5-minute cache write costs 1.25 times the input rate, while a 1-hour write costs 2 times, and a cache read costs 0.1 times.
These rates are based on the assumption that each request reuses the exact prefix and each cache read refreshes the TTL.
For service tiers, Flex is available for certain models like Amazon Nova 2 Lite and OpenAI gpt-oss-120b, but not for Anthropic Claude Sonnet 5 and Opus 5. Flex requests to unsupported models fail with a ValidationException. The Flex rate for Nova 2 Lite is half the Standard rate, as noted in the pricing page.
Reasoning effort settings, such as low, medium, high, xhigh, and max, influence the number of output tokens billed. Lowering the effort can reduce token usage, as demonstrated in testing with a tool-calling agent on Claude Opus 5.
Limitations and availability
Cache settings are model-specific, and the duration choices apply to explicit caching, not implicit caching. For multi-tenant applications, a tenant-specific SHA-256 hash should be prepended to the cacheable prefix. Prompt caching does not work with batch inference, so on-demand inference is recommended.
The supported models and their respective settings can be found in the Amazon Bedrock Pricing page and model cards.
The exact impact of these settings may vary by model and workload, and users should verify the effects using response metadata, logs, and billing data.
