Study finds moderate prompt compression cuts AI inference costs by 28%, but aggressive compression backfires by increasing costs despite token reduction

Pre-registered randomized trial tested prompt compression on Claude Sonnet 4.5 across 358 successful production runs with real orchestration instructions

Moderate compression (50% retention) achieved 27.9% cost savings, while aggressive compression (20% retention) actually increased costs by 1.8% due to longer outputs

Output token prices being several times higher than input tokens means compression economics depend on balancing input reduction against output expansion

Comparison included uncompressed control, three uniform retention rates, and two structure-aware strategies (entropy-adaptive and recency-weighted compression)