The column next to the score
A coding benchmark put three frontier models within eight tenths of a point of each other — and then the column beside the score showed a six-fold spread in cost per task. That gap, and the question of whether anyone outside the labs has reproduced it, runs under most of today: a price cut with a three-month expiry, a demo agent that recommended re-enabling the exact feature a postmortem had disabled, and a state building a complaint registry for data centers.
CursorBench 3.2 results circulating on X — Grok 4.6 at 70.8% and $2.81 per task against Fable 5 Max at 70.5% and $17.32
ARK Invest's…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.