From the source
Amazon's guide explains four request-level Amazon Bedrock settings that change inference cost without changing the model or prompt: cache duration, service tier, reasoning effort, and maxTokens.
- Every Bedrock response carries usage metadata, and input and output tokens bill at the rate for the model and tier that served the request, with cache reads and writes on separate counters and reasoning tokens counted as output.
- For
Anthropic Claude Sonnet 5 and Anthropic Claude Opus 5 on Bedrock, a 5-minute cache write costs 1.25 times the input rate, a 1-hour write costs 2 times, and a cache read costs 0.1 times, which sets the break-even between the two durations.
- A 5-minute cache pays back on the first reuse within 5 minutes, while the 1-hour cache is cheaper from the second request within the hour when the same prefix recurs more than 5 minutes apart.
- Flex is a service tier for latency-tolerant work such as background jobs and nightly summarization; the Amazon Nova 2 Lite and
OpenAI gpt-oss-120b model cards list it as supported, while the cards for Claude Sonnet 5 and Claude Opus 5 list it as not supported.
- A Flex request to an unsupported model fails with a ValidationException rather than silently running on Standard, and Amazon says applications should catch that error and retry once with default rather than retrying every ValidationException.






