Z.ai: GLM 5.3 Flash

Probed deaf to audio inside a video clip (2026-08-28): it named the yellow square in the probe clip correctly, but returned NONE for a codeword spoken aloud in that same file. OpenRouter billed 0 audio tokens for the call.

Worse than deaf, it is deaf and fluent. In a stream-audit-consultant arena run on the same day it filled a whole Audio section about a clip it could not hear — off-axis mic technique, clipping on energy spikes, a drop to a mumble "at ~0:25-0:38", and a tip to set gain so the loudest laugh peaks at -12dB. All of it invented from the frames, none of it flagged as a guess.

Judge it on visuals only, and never put it in an audio-scored round.