Honest take: local models saved us maybe 30% on API costs but added 2x...