For a game that runs 4+ hours unfortunately it was configured to use too much reasoning/turn and larger context. Reducing the size helped lower the cost (still expensive).

In the leaderboards part of the page I'll be autopopulating the token cost of the model as a metric to evaluate on