How the AI Agent Leaderboard works. The whole formula, in plain English.
You should be able to check any number on our board by hand. This page gives you each source, each rule and each limit, including the choices you could reasonably disagree with.
Method
The formula
For each board, we find the model's best configuration, then number the models from 1 upward. Models tied on the source board stay tied.
A model's score is the plain average of its positions on the boards in that list. Lower is better. Example: third, first and third on three boards gives an average of 2.33.
Ties in the average go to the model on more boards, then to the one with the better best rank. If those match too, the models share a place.
Which boards count
The overall list uses one board per publisher, so no one organization gets two votes: Arena Agent, which scores outcomes from real agent sessions; Terminal-Bench 4.0, from Stanford, Harbor and the Laude Institute, with hard tasks in a real terminal; and the Vals Index, which covers agentic finance, coding, legal and tax work.
Category lists go deeper. Coding uses Terminal-Bench plus Vals AI's Code Migration, Vibe Code and ProgramBench. Knowledge work uses Vals AI's Finance Index, Legal Index, Tax Agent Bench and Excel Modeling. Reliability uses four signals from Arena: task outcome, tool calls, shell reliability and steerability. Because those category lists draw on fewer publishers, they are less independent than the overall list.
These boards are not fully separate. The Vals Index includes Terminal-Bench 4.0 as one of its coding components, so results on that test have some influence through two boards. We keep both because the Vals Index covers far more than coding, and we are telling you so you can weigh it.
A board counts only if it tests agent-style work, publishes results openly and has been updated in the past 120 days.
Who gets listed
A model needs a rank on at least two of the three overall boards, or three of the four boards on a category list. One strong result is not enough to appear.
Missing data matters. A model absent from a hard board avoids a result that might have pulled its average down. We show every board's rank, or a dash where there is none, so you can see this for yourself.
The Agreement column
This is the gap between a model's best and worst rank. A gap of 2 or less reads Boards agree, 3 to 6 reads Mostly agree, and anything wider reads Boards disagree. Disagreement is useful: it often means a model is strong at one kind of work and weaker at another.
The weekly update
Every week we re-read each source, recalculate every list and record each model's place, so the Week column shows its movement. The first refresh has no earlier week to compare with, so that column starts empty.
If a source page changes or cannot be read, we keep its last good copy, mark it on the page and fix the reader. A board older than 120 days drops out of the average. If too few boards remain, the update does not publish.
What we do not do
We do not run these benchmarks, test the models ourselves or adjust any result by hand. We do not weight one board above another. And no one can pay for a position. eSEOspace is a member of the Claude Partner Network, and Claude models appear on this list; the formula treats them like any other.
Limits you should know
Averaging positions hides the size of gaps: first and second place could be a fraction of a point apart on the source board or far apart. A list of three boards also moves a lot when one board changes. Both are reasons to read the source boards before making a decision.
Benchmarks can be gamed, and scores from different versions of a test are not comparable. Other respected boards, including Mercor's APEX and Artificial Analysis, are not part of our average right now, and they are worth reading alongside it.
Data credits
Arena data comes from the Arena leaderboard dataset, licensed under CC BY 4.0. We changed it by keeping one configuration per model and renumbering. Terminal-Bench results come from tbench.ai and Vals results from vals.ai. Rights in the underlying results stay with their publishers, and none of them endorse this ranking.
Project Managers who will work with you on your project!
Irina S.hello@eseospace.com
Benjamin G.ben@eseospace.com
David G.david@eseospace.comWork with us
Talk to the team behind your project
Tell us what you need and we’ll match you with the right people.
Book a Discovery Call
Jeanette M.jeanette@eseospace.com
Mark S.mark@eseospace.com
Maryna P.maryna@eseospace.com
Wondering what this costs? Every service has published pricing, and you don't need a discovery call to see it.
View pricingPicking an AI agent for your business?
A benchmark rank tells you which models are strong. It does not tell you which one fits your data, your tools and your risk. Tell us what you want the agent to do and we will give you a straight answer.
Talk to our team
