Stanford HAI benchmarks: best medical AI agents finish under half of real clinical tasks
StanfordHAI · x · 2026-11-13
Two new Stanford HAI benchmarks test medical AI on real clinical and administrative work: while medical AI aces licensing exams, the best agents completed less than half of real-world tasks. HAI faculty affiliates drnigam and jonc101x explain the gap between exam performance and actual job capability — a reality check for medical AI deployment.
More from Models
- Does Qwen3.8-27B survive agent tool-calling loops? Community asks for real-world tests — MajesticPuzzler · 2026-11-13
- Gemini 4 Hydrogen (Flash) spotted rolling out in Antigravity, available to Pro subs — testingcatalog · 2026-11-13
- ChatGPT appears to now render JavaScript when crawling, tests show two very different pages for one URL — gaganghotra_ · 2026-11-13
- OpenRouter data shows token quality uncorrelated with price — priciest GLM 5.3 provider also breaks more tool calls — remilouf · 2026-11-13
- Figure unveils Handoff²: near-frontier accuracy, completes tasks 2x faster — adcock_brett · 2026-11-13
- Speculation mounts that Nvidia turmoil could delay OpenAI's GPT-6.2 launch — VraserX · 2026-11-13