Capital de TokensA 1.8-Point Lead Costs Three Times More
Watch on YouTube A coding agent scores 57.8 percent on CursorBench. Another configuration of the same model scores 56 percent. The first costs an average of 13 dollars and 43 cents per task; the second, 3 dollars and 97 cents. That is a small gap in score and a much larger gap in cost. What are the agents being asked to do? Cursor says
A coding agent scores 57.8 percent on CursorBench. Another configuration of the same model scores 56 percent. The first costs an average of 13 dollars and 43 cents per task; the second, 3 dollars and 97 cents. That is a small gap in score and a much larger gap in cost. What are the agents being asked to do? Cursor says its benchmark evaluates agents on ambiguous tasks involving multiple files, drawn from real Cursor sessions. Its latest version, CursorBench 4.0, adds longer-running problems that ask agents to do more than produce a quick piece of code. And the two configurations? They are Opus 5.5 Max and Opus 5.5 High. Max leads the published table at 57.8 percent. High reaches 56 percent at less than a third of Max’s displayed average cost.
Full episode: https://youtu.be/MMKvbG1hXzE
AI Generated Content: Script, voices and images were produced with AI.
#Shorts