Will the highest score achieved on Humanity’s Last Exam in 2026 be 60% or higher?

Analysis

Humanity's Last Exam serves as a rigorous, multi-modal academic benchmark designed to evaluate frontier-level capabilities across diverse scientific and technical fields. As of late October 2026, the highest recorded performance on this benchmark reached 59.1%, indicating that top-tier systems are currently operating just below the 60% mark. Historical data from BenchLM and other tracking sources show a consistent trajectory of improvement in model performance on complex reasoning tasks. While current results remain slightly under the target, the rapid pace of development in large-scale systems suggests that incremental gains are likely before the end of the year. This assessment incorporates both the current leaderboard standings and undisclosed signals regarding the development cycles of upcoming high-performance models.

Key Evidence

Verified: Scale/CAIS official original-HLE leaderboard top is GPT-6 Astra 54.80 ±1.94 (Sept 17, 2026 update); Fable 5.1 (xhigh) 46.5, Sonnet 5.5 46.1. Polymarket has separate 'HLE-Diamond' markets (GPT Astra 6.1+ Diamond debut ≥60% priced ~85-93%), implying Diamond scores (59.9% CAIS / 60.6% Scale, no tools) do not count for this market. Third-party 59-70% figures (BenchLM, llm-stats, AA text-only 61.4%) are tool-enabled or subset runs and are not the resolution source. Trend on the strict board: 37.5% (Dec 2025) → 44.3% (Mar 2026) → 54.8% (Sept 2026), ~2 pts/month; 65 days remain with Q4 release season.

Risks

No frontier model (GPT-6.1 Astra, Opus 6, Argon Deep Think, Grok 5) is both released and evaluated on the strict full-2,500 no-tools board by Dec 31; GPT-6 Astra's 54.8% stays the top and the Diamond 60%+ scores are ruled a different benchmark — market resolves NO.