Comment by andrewmunsell

The cost-per-task in the charts from the now-remove blog post put it more at Sol-level cost per task, however. It seems like the model is significantly more token efficient in the benchmarks