
Engineering8 min read
Prompt Caching vs Model Downgrade: Which Saves You More?
Caching a 50-turn Sonnet 5.5 agent session cuts input cost 93% ($10.00 → $0.74). Downgrading to Haiku 5.5 costs $0.50 uncached. We ran the numbers — here's which lever wins and when.
Md Badsha · 10 Oct 2026