A lot! We’re actively reducing that though by training our own specialized models. We’re seeing equal or better performance with curated datasets at > 10x cost reduction.