Claude Fable 5.1 leads the AI leaderboards

It depends on the job, which is why Models shows every major leaderboard side by side. The leader of leaders at the top of the home page is the model that finishes highest across all fresh language-model boards (chat, reasoning, coding, agents and vision), scored 10 points for #1 down to 1 point for #10 on each board.

There is no single one. LMArena measures what people prefer in blind votes, Artificial Analysis and Epoch AI run independent evaluation suites, Scale SEAL runs private expert benchmarks such as Humanity's Last Exam, SWE-bench and SWE-Bench Pro test real coding, BFCL and MCP Atlas test tool use. Models tracks them all and shows where they agree.

Each board is reduced to one entry per model (the best variant), ranked, and the top 10 earn 10, 9, 8 ... 1 points. Points are added across the boards counted, and the podium score is points earned over points available, from 0 to 100. Boards not updated in 180 days are shown but not counted. The full method is on the Method page.

Every day at 04:30 UTC, straight from each board's own page, data file or API. Each entry carries the date it was captured, and each board shows the date it last published.

Boards list variants: effort levels such as high or max, thinking modes, dated snapshots, and agent harnesses. Models maps them to one model and keeps the exact name each board published next to every entry.

Yes. /api/summary returns the computed standings as JSON, /llms-full.txt has every board's top 25 in plain text, and every number links back to the board that published it.