A model leaderboard is somebody's choice of benchmarks and weights with a number next to each model. I built Priors because I wanted to be able to change those choices.
You start by putting a few models in the order you think they belong. Priors looks for a mix of benchmarks that gets as close as it can to that order. Then it applies the same mix to the rest of the models with enough data to rank.
You can steer the search toward capabilities, knowledge, problem solving, or a general mix. That tells Priors what kinds of evidence you care about while it tries to match your order.
What you can change
After Priors finds a mix, you can change the benchmark weights yourself. The leaderboard and the index-versus-cost chart move as you adjust them. You can compare reasoning efforts, look at the individual results behind a score, and go back to older data snapshots. When you have an index you like, you can share the link or a card. You can try the builder without signing in.
I also wanted the numbers to be inspectable. Different benchmarks cover different sets of models. A missing result is not a zero, and an estimate is not a measurement. Priors shows where its data came from and lets you require more measured coverage. Set that requirement high enough and some models will leave the ranking. If it can't find a benchmark mix that agrees with your order, it tells you where the disagreements are.
My index
Here's my Capabilities Index. It's the mix I use for the models I care about, and I expect other people to rank them differently. That's why I made the weights editable.
Go to mypriors.com, drag a few models into your order, and see what the benchmarks say about the rest.