ARC-AGI-3 Benchmark Drops and Annihilates Every Frontier Model — Grok Scores Literally Zero

The newest version of the abstract reasoning benchmark designed to test genuine intelligence gives top LLMs less than 1% accuracy while humans score 100%. Grok 4.20 scored 0.00%.

The ARC-AGI-3 benchmark landed this weekend and the results are, as @Hesamation described them, "brutal." Humans score 100% on the test. Every top frontier model scores below 1%. xAI's Grok 4.20 — a model that performs well on standard coding and reasoning benchmarks — scored exactly 0.00%.

Unlock the full briefing

Get every story in today's briefing, the full archive, and the daily AI intelligence brief.

All stories today

Full archive

Daily brief

Cancel anytime. Payments powered by Stripe.