NPCBench is a newly released benchmark for nasopharyngeal carcinoma (NPC), a cancer common in Southern China and Southeast Asia. It tests AI systems on a full care pathway, from initial diagnosis to ongoing treatment decisions, using validated multimodal data—histology, MRI images, and patient's clinical history.

The benchmark is built around 100 patient case histories that contain 417 scored clinical tasks. Those tasks include 100 pathology assessments, 100 MRI T‑stage assessments and 217 care‑decision questions. A key feature is a joint, episode‑level check: a care decision counts as correct only if the diagnosis, MRI stage and treatment choice are all correct for that patient.

Across 30 evaluated models, performance varies widely. In the highest‑scoring multimodal category, a model called Claude‑4.6 achieves a best strict CarePath success of 23 % (i.e., 23 % of patient episodes have all components correct). The same model reaches 89.9 % Care accuracy when diagnosis and treatment are evaluated independently. In contrast, Gemini‑3.1‑Pro‑Preview scores 11.0 % strict success but 89.9 % Care accuracy. A table of these results is shown in Figure 1.