IBM's CEO on the 2026Q3 call: open-weight models now cost "anywhere from 25% to 10% of the token cost of the frontier models" and are "now 90% as effective".
Epoch AI, 16 October: the open cyber frontier sits "roughly where the closed frontier stood 6–7 months ago". GLM-5.3 scores 160 against 171 for Opus 5.5 and GPT-6 Astra.
A tenth of the price for last spring's model. The discount is the lag, and the lag is shrinking. Open weights run on anyone's card. What's missing is a price for the hour on whoever's card it is.
Flop is building that missing piece.