In a keynote in August 2026 I put up four benchmarks I had shown the same audience a year earlier, and three of them were finished. Not failed - finished, scores in the high nineties, quietly retired. I wanted to know whether that was a fluke of the four I happened to pick, so here is every benchmark I could get consistent numbers for, lined up on the same clock, starting from the day each one was published.
The machines behind these scores have their own plot: The Curve follows transistors, training compute, accelerators and model size on one axis.
Scores come from Epoch AI's Benchmarking Hub (CC-BY 4.0), which runs some benchmarks itself under a fixed protocol and collects others from their official leaderboards. Using one collection for everything is what makes the "months to solved" comparison fair, and it is also why a number here can trail a vendor's announcement by a few weeks, or read lower than a leaderboard that uses a different protocol; Humanity's Last Exam, for instance, is scored on the full exam here and on the official Scale AI leaderboard, while text-only subsets score higher. Each panel links its official leaderboard.
The best score line follows the highest result recorded up to each date. "Solved" means that line reached the threshold you pick, and months are counted from the benchmark's publication date; models released before a benchmark were scored on it afterwards, which is why some lines begin left of the gold tick. This page re-fetches its sources and rebuilds itself on a schedule; the stamp at the top is the last run.