Highest score on Humanity's Last Exam before Dec 31, 2026?

HLE scores show variation across reporting leaderboards. As of September 24, 2026, Claude Fable 5.1 is the reported leader on Humanity’s Last Exam (HLE), but specific scores vary significantly depending on the leaderboard version and configuration ^[1]. On LLMBoard/BenchLM, Fable 5.1 registers 65.0%, while Artificial Analysis lists it at 59.1% ^[1]. Scale Labs’ finalized 2,500-question leaderboard reports GPT-6 Astra at 54.80% and Fable 5.1 (xhigh) at 46.50% ^[2]. Scale Labs notes its finalized 2,500-question HLE is distinct from prior versions, making direct score comparisons difficult without identifying the benchmark release and evaluation protocol ^[2]. On a widely referenced group of leaderboards, Claude Fable 5.1 is listed at 59.1%, ahead of Claude Fable 5 at 55.5% and Claude Opus 5 at 54.9%; several aggregators confirmed this ordering in mid-to-late September 2026 ^[3]. Earlier HLE results were lower, with X-Masters (SciMaster) achieving 32.1% in July 2025, which was presented as the first result above 30% ^[4]. The original HLE, released in January 2025, listed o3-mini-high at 13.4% on its text-only subset ^[5]. HLE-Verified, updated in February 2026, increased reported model accuracy by 7–10 percentage points generally, and 30–40 points on specific items ^[6].

Prediction markets indicate a high probability for 50%+ by year-end. Kalshi tracks achievement thresholds ranging from 50% to 75% for resolution by December 31, 2026 ^[7]. Polymarket’s contract specifies a "Yes" resolution if any model reaches at least 60% on the official HLE leaderboard (agi.safe.ai) by December 31, 2026, 11:59 PM ET ^[8]. Market evidence strongly suggests a score of at least 50% will be reached by year-end; a Kalshi odds mirror reports 99% implied probability for ≥50%, and Octagon shows 87% market probability for ≥50% versus 7% for ≥90% ^[7]. Polymarket’s displayed market state for the broader 2026 outcome is roughly 70% ^[8]. The best forecast for the highest official score by December 31, 2026, is approximately 60–65%, with a central estimate of 62% ^[1]. This forecast is based on the current 59.1–65.0% reported range, anticipated frontier-model releases, and the structure of contract thresholds ^[1]. An earlier analyst forecast, not an observed prediction-market price, set a central estimate of 50% for the highest score by December 31, 2026, with a plausible range of 40–60%, expecting a frontier tool-augmented reasoning system to lead ^[5].

Model scaling drives HLE gains, but benchmark changes add uncertainty. Experts attribute recent performance gains to scaling, stronger reasoning methods, architectural innovation, and competition among key developers such as Anthropic, OpenAI, and Google DeepMind ^[8]. Major uncertainties include new model releases, changes to evaluation protocols, and revisions to HLE-Rolling or overall benchmark quality ^[8]. A July 2026 psychometric analysis found that HLE’s eight subject labels might not represent distinct latent abilities, as a general factor accounted for 98–99% of item-response variance in the multiple-choice subset ^[9]. This suggests caution in interpreting domain-specific rankings ^[9]. Expert opinion indicates HLE scores above 50% would signal expert-level closed-ended academic question answering, though not autonomous research ability or AGI; benchmark validity and item quality remain central uncertainties ^[5]. No retrieved source established an official prediction-market contract specifically resolving the highest HLE score by December 31, 2026 ^[5]. Adjacent prediction market research, such as TruthTensor’s study (December 2025–January 2026) and Prediction Arena’s findings on model performance in different market formats, suggests caution when translating AI benchmark capabilities into market probabilities ^[10].