AWS publishes cost-per-task and agent ROI framework for Bedrock

AWS's Machine Learning Blog published a framework for judging Amazon Bedrock spending by cost per task rather than cost per token, arguing that tokenizer efficiency, output verbosity, agentic loop count and reasoning overhead can make a lower-priced model cost more in practice. In the post's single worked example, a model with twice the list price finished 65% cheaper per task; the post also recommends caching, routing and output constraints to cut cost per task, and proposes measuring agent ROI against human baselines with holdout experiments and a five-dimension scorecard. A follow-up AWS post details four Bedrock inference settings that change cost without changing the model or prompt: cache duration, service tier, reasoning effort and maxTokens. It states that on Claude Sonnet 5 and Claude Opus 5 a 5-minute cache write costs 1.25x the input rate and a 1-hour write 2x, while cache reads cost 0.1x, and that Flex tier requests to unsupported models fail with a ValidationException. These are the posts' own recommendations and figures, not independent benchmarks.

Generated from reports · updated 17 h ago

1. November 13 04:51 · 1 reports
   Four Amazon Bedrock inference choices that affect cost
   AWS Machine Learning Blog：Four Amazon Bedrock inference choices that affect cost
2. November 13 02:43 · 1 reports
   Publishing a framework for cost per task and agent ROI on Be
   AWS Machine Learning Blog：Optimizing cost per task and proving agent ROI on Amazon Bedrock

Not enough continuous observations yet to draw a trend.