Abstract: To investigate the strength performance of aligned steel fiber-reinforced geopolymer composites (ASFRGPC), specimens with varying geopolymer mix proportions and steel fiber aspect ratios were designed. Tests involving cube compressive strength, prism axial compressive strength, cube splitting tensile strength, and prism flexural strength were conducted. On this ground, the failure mechanisms, strength variation patterns, and interrelations between different strengths were explored and strength characterization models were established. The results show that variations in mix proportions and material compositions have a small impact on the interrelationships between different strengths of plain geopolymer. The prism and cube compressive strengths exhibit an approximately linear relationship, with the proportional coefficient closely matching those reported in the literature. The splitting tensile strength and flexural strength demonstrate a power-law function relationship with prism compressive strength. The model recommended in Australian standard for geopolymer and alkali-activated concrete structures can directly represent these relationships, with an average absolute prediction error of less than 12%. Aligned steel fibers can significantly improve the tensile and flexural performance of ASFRGPC. Compared to plain geopolymer, the splitting tensile strength and flexural strength of ASFRGPC increased by at least 120% and 100%, respectively. In contrast, randomly distributed steel fibers typically enhanced these strengths by no more than 50% and 80%, respectively. Aligned steel fibers affect compressive strength as well. They enhance the compressive strength perpendicular to the axial load while reduce the strength parallel to the axial load. The strength model for ASFRGPC can be established as a combination of the strength model of plain geopolymer and a linear function of fiber characteristic value. The resulting models for splitting tensile strength, flexural strength, and elastic modulus achieve excellent predictive performance, with an average absolute prediction error not exceeding 5%.