Microsoft released SDBench, a benchmark for sequential medical diagnosis. It tests AI systems that gather information, order tests, and commit diagnoses step by step. A gatekeeper returns real or synthetic test results, and a judge scores diagnoses on a 1–5 scale. The benchmark ships with a small synthetic dataset for development. Clinically sourced datasets require separate access. Configurations pin gatekeeper and judge models for comparability. Costing modules estimate monetary test costs. The Science paper configuration reproduces published baselines. A website hosts full documentation. Local, Azure, and OpenRouter model routing are supported.