Right, the discount does not matter if the workload itself is too token-hungry for hosted capacity to stay stable. Multi-agent coding just makes the meter hurt faster. That's a big part of why we're building rig.ai around local inference instead.