The cause isn't clinical knowledge, it's agentic behavior. In ~60% of turns, Opus 5 hallucinates the API's response inside its own message (emitting a tool call and a fabricated server reply together), instead of issuing one call and waiting. Opus 4.8 did this ~7% of the time. To isolate the effect, we ran a diagnostic that ignores the fabricated text and executes only the real call. Score rises to ~68% but still short of 4.8. Retrieval tasks recover; multi-step order-writing tasks stay low (potassium replacement 7%, HbA1c ordering 23%), and it sometimes reports real patients as "not found." A capable model held back by not respecting the tool-use loop. Note: the ~40% is the leaderboard figure, same harness as every other model. The ~68% is a diagnostic only, with a temporary parser tweak, and isn't directly comparable. 📊 Leaderboards: medicalsphere.ai/benchmarks 🧪 Full pipeline: github.com/medicalsphere/Med…