Bivalent peptide toxins comprising two cysteine-rich domains have evolved from single-domain precursors on multiple occasions in animal venoms, resulting in enhanced molecular target selectivity and avidity. Although bivalent toxins are emerging as prevalent in animal venoms, the genomic and evolutionary processes driving the transitions between single- and multi-domain architectures remain poorly understood. Here, we investigated the evolution of bivalent inhibitor cystine knot (ICK) toxins in spider venom. We first generated generate a genome assembly of the tree-dwelling funnel-web spider Hadronyche cerberea, revealing a massive expansion of ICK toxin-encoding genes, including the bivalent π-hexatoxin-Hc1a. All ICK toxin genes share a conserved three-exon-structure, flanked by transposable elements (TEs) that may have facilitated gene expansion. This gene structure is shared by the Hc1a family, where the entire mature bivalent toxin is encoded by the third exon. Leveraging de novo transcriptome assemblies of 86 spider species along with venom proteomic data, we show that bivalency in the Hc1a family is of ancient origin, but with recurrent domain expansions and losses due to point mutations, deletions, and unequal crossing-over facilitated by high interdomain sequence similarity. In contrast, the bivalent toxin DkTx from Cyriopagopus schmidti appears to have evolved once in a restricted group of tarantulas. While we also find here multiple domain loss events, these are potentially facilitated by TEs as opposed to pseudogenisation or unequal crossing over. Our findings reveal that singular events of domain duplication can give rise to complex, asymmetrical evolutionary trajectories shaped by gene instability and selective retention of functional domains.