HalachaBench: leaderboard

HalachaBench (v1 early results, provisional) asks models to answer real questions from the Chief Rabbinate of Israel's semicha (rabbinic ordination) exams: the 60 questions of five papers of the Cheshvan 5787 sitting, each answered three times in Hebrew, closed-book and without tools. One AI judge, GPT-6 Astra, checks every answer against a sealed rubric of 9 to 12 criteria that AI models wrote and reviewed; an answer's score is the share of rubric points it earns. Unlike LEXam, where law-exam answers are graded against the examiners' own reference answers, these exams have no official answer key, so the grading rests on AI-written rubrics that no rabbi has reviewed yet. The judge is also a contestant. Open-weight models that ran as local 4-bit files of higher-precision releases are left off the board. This board is the mean score, from 0 to 100, over a model's 180 answers.

Metric: Mean rubric score on 60 semicha exam questions, three attempts each, GPT-6 Astra judge (0-100). Source: github.com. Saturation forecast: Around February 2027. 23 models tracked.

Top models

#ModelScoreOverall rank
1Claude Fable 5.2 (High)85.77#1 (Claude Fable 5.2)
2Claude Fable 5.1 (High)79.74#14 (Claude Fable 5.1)
3Gemini 4 Argon (High)78.22#11 (Gemini 4 Argon)
4GPT-5.6 Sol (High)77.27#29 (GPT-5.6 Sol)
5GPT-6 Astra (High)73.32#6 (GPT-6 Astra)
6GPT-6.1 Sol (High)72.76#10 (GPT-6.1 Sol)
7Claude Opus 5.5 (High)69.05#5 (Claude Opus 5.5)
8Gemini 3.8 Flash (High)64.37#36 (Gemini 3.8 Flash)
9GPT-6 Sol (High)48.67#24 (GPT-6 Sol)
10Muse Wonder (High)48.23#27 (Muse Wonder)
11Claude Opus 4.8 (High)48.02#56 (Claude Opus 4.8)
12Muse Spark 1.3 (High)36.23#42 (Muse Spark 1.3)
13GPT-6.1 Terra (High)34.6#19 (GPT-6.1 Terra)
14GPT-6 Luna (High)32.37#93 (GPT-6 Luna)
15Grok 4.7 (High)29.87#43 (Grok 4.7)

No result here: #2 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=halachabench · How It Works · Data refreshed daily, snapshot 2026-11-12.