It is how many calls are bundled into one AI request. Larger batches cost less because the rubric is sent once per request instead of once per call. The trade-off is that a failed request loses every call in it, and a model shown several calls together tends to score them against each other rather than against the rubric.