This research was conducted in order to test the effectiveness of Problem-Based Blended Learning in improving students' collaboration skills. The purpose of this study is to describe the significance of learning differences between students who are taught by using Problem-Based Blended Learning model (BL-PBL), PBL model, conventional models, and the effectiveness of these learning models. The type of this research was quasi-experimental research, by conducting pre-test and post-test control group design. The subjects of this study were the eleventh grade students of SMAN 1 Sleman, while the number of the students in experimental classes was 33 students, and 35 students belong to the control classes (with a total number of 101 students). Experimental class I was taught by using Problem-Based Blended Learning, while experimental II was taught by using PBL learning model, and the control class was taught conventionally. The data of students' collaboration skills were obtained from the result of the questionnaire sheets during pre-test and post-test and by using observation sheets (filled out by observers) in each learning process. The data were analyzed by using N-Gain, and the effectiveness of students' collaboration skills showed by the value of the effect size calculated by using Partial Eta Squared. The results of this study indicate that the value of the effect size on the BL-PBL model was 0.773, and categorized as high (compared to other models). In addition, the N-gain value also showed that BL-PBL learning was in the high category (0.4). Hence, it can be concluded that BL-PBL learning is more effective in increasing students' collaboration skills rather than PBL and conventional models.