Model comparison
Choose two to four models. Every number comes from the same source on the same scale, so there is no v4-versus-v5 trap.
How to use it
Eliminate first, then compare
If the context is too small, the licence does not fit, or it is too slow — cut it immediately. Compare scores only among the two or three that survive.
Watch blended price, not unit price
A model that is cheap on input and expensive on output can cost more than a seemingly pricier one when answers are long. A 3:1 weighting is a reasonable approximation for most workloads.
Calibrate on cost per task
Two models at $2/$10 can differ 3× in output tokens. The cost-per-task figure in the table comes from AA's measurements, which tracks the real bill far better than the unit price.