ArmBench-LLM v2 - Overall: leaderboard

ArmBench-LLM 5-category edition (Metric AI Lab, Yerevan; 2026-11-20 reorganization): Armenian-language benchmark run with Metric's lighteval fork, regrouping the 24 tasks into five equal-weight categories (Knowledge, Reading, Analysis, Generation, Correction). Unweighted mean of the five category scores (BLEU/100, national exams / 20), 0-100, higher is better. The site excludes Claude 4.5 and 4.6 models, citing evaluation reliability issues.

Metric: Mean of 5 category scores (0-100; BLEU/100, exams/20). Source: huggingface.co. Saturation forecast: Around December 2027. 32 models tracked. Updated 22 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-6.1 Sol65.14#15
2DeepSeek V4.1 Flash62.86#69
3Ayb-0.5-9B61.64
4GPT-5.2 Pro60.17#139
5Gemini 3 Flash (Preview)59.51#156
6Claude 3.7 Sonnet59.1#455
7Gemma 4 31B (IT)58.71#354
8Ayb-0.5-4B58.28
9Grok 4 Fast57.88#446
10Gemini 2.5 Flash57.35#466
11Claude Sonnet 457.29#414
12Gemini 3 Pro (Preview)56.65#123
13GPT-5.4 Mini56.12#409
14Qwen 3.5 27B55.03#380
15Grok 4.1 Fast54.6#431

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=armbench-llm-v2-overall · How It Works · Data refreshed daily, snapshot 2026-11-22.