Show HN: Does per-step reasoning effort save money in Claude Code? (github.com)

A new research initiative has explored the cost-saving potential of dynamic reasoning effort selection in Claude Code, demonstrating that allowing the decision model Jev to adjust reasoning effort on a per-step basis significantly reduces costs. At maximum reasoning effort, testing revealed an impressive 55% reduction in costs while maintaining success across all tests. In longer sessions, estimated savings ranged from 4% to 9%, highlighting that the majority of expenses stemmed from re-reading context rather than thinking effort alone. This development is significant for the AI/ML community as it showcases a novel approach to optimizing AI resource usage without sacrificing performance. The study's findings indicate that in high-effort scenarios, there’s limited room for savings since routine steps require minimal computation. However, the ability to dynamically adjust reasoning effort during sessions promises better resource management and efficiency in larger tasks, capitalizing on the fact that many agent steps can be handled at lower computational costs. With a high cache hit rate and effective handling of effort adjustments, the framework opens avenues for further exploration in AI performance optimization and cost-efficient deployments.