GPT-5.6 Sol (high) vs GLM-5

GPT-5.6 takes this one, and the Elo intervals don't even touch, so I'm comfortable calling it. The gap that actually matters is number handling: our anti-slop and numbers check puts GPT-5.6 at 91.0 and GLM-5 at 77.3, the widest split between these two, and GLM's variance there is wild. Some GLM scripts reframe stats the way a human would, others recite them like a quarterly report. Structure is the second miss, a gap of about seven and a half points on YouTube best practices, which in practice means rebuilding retention beats yourself. Now, credit where it's due: GLM-5 is a real open-weights model, its hooks land nearly as hard as GPT's, and at an estimated cent and a half per script versus fifteen cents, it's roughly a tenth of the cost. If you're generating at volume and editing anyway, that math is tempting. One caveat I have to flag: GPT-5.6 sits on our judge panel and scores itself generously. The other two judges still prefer it, just by less. For a publishable draft, GPT-5.6.

Pick GPT-5.6 Sol if you want the closest thing to a publishable draft and can live with fifteen cents per script.
Pick GLM-5 if you want open weights at roughly a tenth of the cost and don't mind rewriting the stats-heavy stretches yourself.