REDDIT · REDDIT// 27d agoBENCHMARK RESULT

AI tops 40% on Humanity's Last Exam

Humanity's Last Exam (HLE), a 2,500-question benchmark co-developed by the Center for AI Safety and Scale AI and published in Nature, now sees top models scoring ~40%—up from single digits at launch in early 2025. Expert humans in their domains still average ~90%, making the gap stark.

// ANALYSIS

The jump from sub-10% to 40%+ in roughly one year is remarkable, but the remaining 60% gap shows frontier AI still lacks deep expert-level reasoning.

–HLE was built by nearly 1,000 researchers who deliberately excluded any question current AI could answer, making it a genuine frontier benchmark by design
–At launch: GPT-4o at 2.7%, o1 at 8%; by March 2026, Gemini 3.1 Pro Preview leads at 44.7%, Claude Opus 4.6 Thinking at 34.4%
–Expert humans in their respective fields average ~90%—a 45+ point gap that persists despite rapid AI improvement
–Now published in Nature, HLE has become the canonical academic benchmark for measuring frontier model progress
–The rapid improvement rate raises an uncomfortable question: how long before HLE itself needs to be replaced?

// TAGS

benchmarkllmreasoningresearchsafety

DISCOVERED

27d ago

2026-03-16

PUBLISHED

27d ago

2026-03-16

RELEVANCE

8/ 10

AUTHOR

PixeledPathogen