An open-weight model gets the same score as closed models at a low cost

The authors write that this is correct only when the model has sufficient turns. The limit on messages gives incorrect low scores.