Skip to content

Humanity's Last Exam

Best independently measured score without tools (Scale AI leaderboard), by model release date.

Topic: Benchmarks

What this measures: 2,500 expert-written questions chosen because the best AI of early 2025 got them wrong. At launch, top models scored under 10%.

0%50%100%20252026

Scale has listed no model released after September. Artificial Analysis's independent run, on the 2,158 questions without images, puts Claude Fable 5.2 at 74.3%. It is not on this line because the two testers can score the same model up to 13 points apart.

Labs report higher scores with tools (e.g. 64.7% for Claude Mythos Preview, self-reported). Those are not comparable with these no-tools runs. Reviews by FutureHouse and the exam's own team found problems with 18–29% of the biology and chemistry answers, so scores above about 70–80% may reflect errors in the answer key. The cleaned HLE-Diamond subset (1,000 questions) is scored separately.

Sources: Scale AI, Artificial Analysis, CAIS / Scale AI, FutureHouseData as of 12 Nov 2026
Show data table and sources
DateLabelValue
17 Dec 2024o18%
16 Apr 2025o3 (high)20%
5 Jun 2025Gemini 2.5 Pro22%
7 Aug 2025GPT-525%
6 Oct 2025GPT-5 Pro32%
18 Nov 2025Gemini 3 Pro38%
19 Feb 2026Gemini 3.1 Pro (high)46%
3 Sept 2026GPT-6 Astra55%

Entries with this chart