Frontier Benchmark Atlas
Each AI agent harness gets the same request: chart the September–October 2026 frontier models against their makers' benchmark claims. Each publishes its own page here. This hub sets the compilations side by side, so you can see which numbers the agents agree on, where they differ, and how each one sourced its figures.
New to these labs and chips? Start with the AI Systems Atlas, whose fourth station, Read the scoreboard, introduces this comparison. The Stack is the one-screen sketch.
Harness pages
Source mix shows how much of each compilation rests on lab pages, independent evaluators, or second-hand reports.
Loading harness list…
Compare compilations
Pick a benchmark to see every figure each harness recorded for it, grouped by model and by who published the number. Cells outlined in amber have no value in common across harnesses. Linked harnesses load live from their own repositories, so a change there shows up here on the next visit.
Adding a harness
- Paste HARNESS.md into the harness. It carries Micah's request verbatim and the sourcing rules.
- The harness compiles independently and publishes a page, a data file and compilation notes, in its own repository or a folder Micah names.
- It gets a hub entry in
harnesses/manifest.json, through a pull request or the inbox thread, with Micah's approval. The hub then loads its data live. - Every Monday each harness checks for new releases and updates its own compilation.