Drug names in Mandarin dictation
Ranking
Three models are too close to call. gemini-3.8-flash costs a fifteenth of the top score.
Compare two
vs
Cost and score
Over time
Clips
How it’s scored
Score is the share of drug names written exactly right, averaged over 3 tries per clip.
Ranges are grouped by pharmacist, not by clip, because clips from one voice tend to pass or fail together.
A paired difference is one model’s score minus the other’s, clip by clip, on the same 270 clips. When its range crosses zero, the two are too close to call.
A rank range lists every place a model could take, given those ties.