You'll get 50% prediction accuracy for a fair coin no matter what result a model gives. Try testing adversarial games. Mixed strategies make actual sense in them.
The model gives a probability distribution as an answer; you can check to see if two arbitrarily chosen probability distributions are similar or not.
I mean predicting the result of a fair coin toss is not a classification task those models are trained to do. It can be framed as a decision making task, but for a fair coin any probability distribution of heads/tails will do equally well.
BTW, for a task of predicting a 51/49 biased coin, the correct answer is 100/0.
I'd say selecting a scoring rule that depends on the probability distribution (the Brier score, for example) changes the task from predicting the results of coin flips to predicting the distribution. Maybe it will work with a corresponding prompt.