LiveBench Reasoning

In plain words · How many freshly written logic puzzles it solves each month

What does it measure?

Skill at logic puzzles: drawing conclusions from several people's true/false statements, placement puzzles with interlocking constraints, spatial reasoning about shapes — thinking things through rather than knowing facts.

Who evaluates it, and how?

LiveBench, built by researchers at Abacus.AI, NYU and elsewhere, releases new questions every month from recent competitions, papers and news, reducing the chance that an AI has seen the questions during training (contamination). Only questions with fixed answers are used, graded by program rather than by another AI. This table takes the latest monthly results published at livebench.ai; an AI missing from the newest release is listed separately below the table with its last published score, unranked. The reasoning subject consists of truth/lie deduction (Web of Lies), Zebra puzzles and spatial reasoning tasks.

Reading the score

0–100. The questions are designed to be hard — even top AIs struggle to exceed 70 — and because they change monthly, avoid comparing scores from different dates. It measures pure logic, not common sense or knowledge.

Last updated: 2026-06-25