A pair of Nature Biotechnology reports introduced DeepSomatic, a deep‑learning somatic small‑variant caller compatible with short‑ and long‑read sequencing, and a set of public CASTLE benchmark panels derived from cancer cell lines. The tools and datasets aim to address persistent gaps in somatic variant detection and cross‑platform benchmarking. DeepSomatic demonstrated improved accuracy across technologies by harmonizing model architectures and training on diverse data modalities; CASTLE provides seven benchmarking datasets to enable reproducible comparisons and training. Authors dispelled the trade‑offs between read length and caller performance by offering methods tailored to each platform while preserving cross‑compatibility. The work facilitates more reliable tumor profiling for clinical genomics, enabling labs to select callers and platforms with clearer performance expectations and helping drug developers and diagnostics companies standardize variant calling pipelines for translational and regulatory use.