NoteEvaluation & Benchmarks

Same task, six rounds: 3.6 times the time, 15 times the cost

A record of putting exactly the same prompt six times into two commercial models released on the same day, having them make 3D games, motion graphics and trailer clips, and then comparing them by operating the results directly on screen. The quality of the outputs tilts to one side, but the part worth noting is the price revealed at the end: the time spent on the same six tasks is given as 4 hours vs. 14 and a half hours, and, converted into fees, $49 vs. $600. This means a quality edge drives up costs by a multiple, and the numbers make clear why a design that sets the best model as the default for jobs that run hundreds of times a day cannot hold up. As a conclusion, it offers a division of labor: design and judgment go to the expensive one, repeated execution to the cheap, fast one. However, all six tasks fall within a single genre, 3D graphics generation, the judging rests only on visual impressions, and how the two figures were calculated is not disclosed.

Key points

  • Six head-to-head rounds matched on the same prompt, the same settings and the same external tools
  • Total time 4 hours vs. 14 and a half hours; converted into fees, $49 vs. $600
  • Recommends a division of labor: judgment to the expensive model, repeated execution to the cheap one
  • All the tasks were built from a blank slate, so the ability to modify existing code was not checked

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.

Same week's notes

Is it still holding the first rule at the tenth file?Agents meant to be isolated exchanged signals with each otherA design that doesn't hand regulatory review wholesale to one modelWhat held back the switch was accumulated setup, not performanceA flood of malicious packages in a short span, and the bar for isolation