LetTokenBurn
Sign in
Verified100%Harbor nativeAbstract Reasoning

ARC-AGI-2

Abstract Reasoning Corpus v2

ARC-AGI-2: 更难的抽象推理

PaperRepoDatasetLeaderboardHomepage

How to run

ltb · arc-agi-2
$ npm i -g lettokenburn-cli
$ ltb bench run arcprize/arc-agi-2@1 --benchmark arc-agi-2 -a claude-code -m anthropic/claude-opus-4-8
$ ltb bench push

Harbor-native: you can also call Harbor directly.

harbor
$ harbor run -d arcprize/arc-agi-2@1 -a claude-code -m anthropic/claude-opus-4-8

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

118 sourced entries
HarnessSourceConfidenceLink
GPT-6 Astra95.0N/AIndependentlowN/Aepoch.ai
GPT-6.1 Sol94.2N/AIndependentlowN/Aepoch.ai
Claude Opus 5.592.5N/AIndependentlowN/Aepoch.ai
GPT-5.6 Sol92.5N/AIndependentlowN/Aepoch.ai
Claude Opus 590.4N/AIndependentlowN/Aepoch.ai
Claude Fable 5.190.0N/AIndependentlowN/Aepoch.ai
GPT-6 Sol89.6N/AIndependentlowN/Aepoch.ai
Claude Fable 589.2N/AIndependentlowN/Aepoch.ai
Gemini 3.8 Flash89.2N/AIndependentlowN/Aepoch.ai
GPT-5.585.0N/AIndependentlowN/Aepoch.ai
Gemini 3 Pro84.6N/ASelf-reportedlowJun 1, 2026arcprize.org
Gemini 3.7 Flash84.6N/AIndependentlowN/Aepoch.ai
Gemini 3 Deep Think84.6N/AIndependentlowN/Aepoch.ai
GPT-5.5 Pro84.6N/AIndependentlowN/Aepoch.ai
GPT-5.6-terra83.9N/AIndependentlowN/Aepoch.ai
GPT-5.4 Pro83.3N/AIndependentlowN/Aepoch.ai
GPT-5.4 Pro83.3N/ASelf-reportedlowJun 1, 2026arcprize.org
Gemini 3.1 Pro77.1N/AIndependentlowN/Aepoch.ai
Claude Opus 4.775.8N/AIndependentlowN/Aepoch.ai
GPT-5.474.0N/AIndependentlowN/Aepoch.ai
Claude Opus 4.872.1N/AIndependentlowN/Aepoch.ai
Gemini-3.5-flash72.1N/AIndependentlowN/Aepoch.ai
claude-opus-4-6-120K69.2N/AIndependentlowN/Aepoch.ai
Claude Opus 4.668.8N/ASelf-reportedlowJun 1, 2026arcprize.org
Grok-4.667.1N/AIndependentlowN/Aepoch.ai
GLM-5.3 Flash65.8N/AIndependentlowN/Aepoch.ai
Grok 4.2065.1N/AIndependentlowN/Aepoch.ai
DeepSeek V4 Flash 073161.4N/AIndependentlowN/Aepoch.ai
DeepSeek V4 Pro 081361.3N/AIndependentlowN/Aepoch.ai
Claude Sonnet 4.660.4N/AIndependentlowN/Aepoch.ai
Gemini 3.6 Flash60.4N/AIndependentlowN/Aepoch.ai
Kimi K360.4N/AIndependentlowN/Aepoch.ai
GPT-5.6 Luna59.5N/AIndependentlowN/Aepoch.ai
GPT-6 Luna59.3N/AIndependentlowN/Aepoch.ai
Claude Sonnet 4.658.3N/ASelf-reportedlowJun 1, 2026arcprize.org
GPT-5.2 Pro (2025-12-11)54.2N/AIndependentlowN/Aepoch.ai
GPT-5.2-chat52.9N/AIndependentlowN/Aepoch.ai
Grok 4.552.6N/AIndependentlowN/Aepoch.ai
Qwen3.8 27B42.4N/AIndependentlowN/Aepoch.ai
Inkling Small40.1N/AIndependentlowN/Aepoch.ai
claude-opus-4-5-20251101-64K37.6N/AIndependentlowN/Aepoch.ai
Inkling36.5N/AIndependentlowN/Aepoch.ai
Gemini 3 Flash Preview33.6N/AIndependentlowN/Aepoch.ai
Gemini 3 Pro31.1N/AIndependentlowN/Aepoch.ai
Gemini 3.1 Pro31.1N/ASelf-reportedlowJun 1, 2026arcprize.org
claude-opus-4-5-20251101-16K22.8N/AIndependentlowN/Aepoch.ai
GLM-5.222.8N/AIndependentlowN/Aepoch.ai
GPT-5.4-mini18.9N/AIndependentlowN/Aepoch.ai
GPT-5 Pro (2025-10-06)18.3N/AIndependentlowN/Aepoch.ai
GPT-5.117.6N/AIndependentlowN/Aepoch.ai
Top 50 of 118 by score

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

ARC-AGI-2 (Abstract Reasoning Corpus v2) — 更难的抽象推理. Scoring: 准确率. Difficulty: 极限. Imported from benchmark_research.md (2026-06).