Comment by jiggawatts

A.k.a.: Too expensive for inference at fp16 and too dumb after quantisation, so it wouldn’t look good to release a step backwards.

GPT 4.5 was scrapped for similar reasons.