Compare evaluations

What each lab measured, how deeply, and how that shifted between October 2025 and October 2026. All cells come from the documents themselves. Click any cell or point to open the source at the cited page.

Safety evaluation coverage by document

Safety breadth over time

The number of safety areas (out of 21) that each document covers at “light” depth or deeper. There is one panel per lab, and each point is a document.

Benchmark explorer

Choose a benchmark to see every score reported for it. Scores are as published: harnesses, settings and thinking budgets differ across labs, so the numbers are not strictly comparable.

Which benchmarks get reported, and for how long

These are the benchmarks reported in at least 3 documents, with the first and last document that reported each one. A short span or a single lab often means a benchmark was saturated or replaced.