Gene fusions are common drivers and therapeutic targets in cancers, but clinical-grade bioinformatics callers are lacking. Here we introduce a novel method SplitFusion, which is fast by leveraging BWA-MEM split alignments, can detect cryptic splice site fusions, and can infer frame-ness and exon-boundary alignments for functional prediction and minimizing false-positives. SplitFusion demonstrates superior sensitivity, specificity, accuracy and consumes minimal computing resources. In our study of 1,076 formalin-fixed paraffin-embedded lung cancer samples, SplitFusion detected not only common fusions (EML4 4.7%, ROS1 2.0% and RET 1.1%) with various partners, but also rare (KLC1-ALK, CD74-NRG1, and TPR-NTRK1) and novel (FGFR3-JAKMP1, CLIP2-BRAF, and ITPR2-ETV6) fusions. In 35 glioblastoma samples, SplitFusion-Target detected six (17%) EGFR vIII (exons 2-7 deletion) cases. Furthermore, we find that the EML4-ALK variant 3 is significantly associated with occurrence of multiple breakpoint-defined subclones, namely high intratumor heterogeneity. In conclusion, SplitFusion is well-suited for clinical use and for studying fusion-defined tumor heterogeneity.