We describe a new branch predictor that is designed to balance multiple constraints—predicting branch biases versus predicting specific branch instance behavior. Most branch instances only require branch bias information for accurate predictions while a select few require more sophisticated prediction structures. Our predictor uses a cache mechanism to classify branches and dynamically adjust the balance of the predictor. On average, our predictor mispredicts 24% less often than YAGS and 19% less often than a global perceptron predictor with the same bit budget.