MAI-DxO reaches Science, opens SDBench

The important advance is orchestration, not another medical model: structured debate, evidence gathering, and cost-aware test selection can materially improve frontier-model performance.

• –SDBench evaluates the diagnostic path—not just the final answer—by scoring questions, tests, accuracy, and estimated cost. SDBench
• –The open-source harness gives developers a practical way to compare models under interactive clinical workflows.
• –NEJM and MIMIC datasets remain gated, while only synthetic cases are fully access-free, limiting immediate reproducibility.
• –Follow-up testing on newer models such as Gemini Argon suggests the orchestration layer may generalize beyond a single model generation.
• –The headline results are still benchmark results, not evidence of safe deployment; Microsoft’s planned hospital study is the more consequential test.