Even when neither the model nor the prompt changes, inference bills on Amazon Bedrock can keep climbing. AWS is warning enterprises to pay attention to four request-level choices: cache duration, service tier, reasoning effort, and maximum output tokens. Cloud platforms are turning AI cost control from an R&D problem into configurable operational work.
What this is
Amazon Bedrock is AWS's enterprise service for accessing large language models. This test focuses on four parameters that do not change the model but do affect billing: cache duration, service tier, reasoning investment, and maxTokens (the maximum number of tokens allowed in a single response).
Prompt caching can reuse recurring system instructions, tool definitions, and reference materials. AWS now offers explicit cache durations of five minutes and one hour, but a longer duration is not necessarily cheaper. With Claude Sonnet 5 and Claude Opus 5, for example, five-minute cache writes cost 1.25 times the standard input price, one-hour writes cost twice the standard input price, and cache reads cost only 0.1 times the standard input price. Enterprises need to calculate the break-even point based on reuse frequency.
Reasoning tokens—the computation models use to analyze a problem rather than directly produce the answer—are billed as output tokens. Increasing reasoning investment or the maximum output limit can also quietly expand costs in the background. AWS recommends checking the actual impact through response metadata, logs, and billing data.
Industry view
We see the point of this piece as less another model upgrade than a cost dashboard that enterprises must add when calling AI. Model price tags are becoming increasingly easy to compare, but what ultimately determines the monthly bill is often cache hit rate, service tier, task complexity, and output limits.
The risks are equally clear: caching does not hit on every request. Bedrock's implicit caching is best-effort and does not guarantee reuse, while supported durations and parameters vary across models. If enterprises select tiers solely by theoretical pricing without recalculating costs from real request logs, cost optimization can easily become another billing mystery.
For cloud vendors, as model capabilities converge, call orchestration, monitoring, access control, and cost governance will become key drivers of renewals and expansion. This also means an AI product's gross margin may increasingly depend on how it is used and how well customers are managed, not only on the scale of its model parameters.
Impact on regular people
For enterprise IT: When procuring large model services, enterprises should assess cache reuse rates, the cost per call at each tier, and reasoning tokens in one unified evaluation, rather than comparing only model unit prices and benchmark results.
For workplace AI users: Enterprise AI tools will increasingly be tiered by task instead of defaulting to the highest specification. At the same time, boundaries, budgets, and audit requirements for employee use of AI will become clearer.
For the consumer market: Products with agents may adopt different service tiers based on complexity. Some high-frequency scenarios will use caching to reduce costs, meaning users may no longer need to bear the full cost of inference for every simple request.