I also notice this with plus account. Over-optimization of inference to reduce costs/increase scale maybe