Nejumi 4 - ARC-AGI-2: leaderboard
ARC-AGI-2 abstract grid-reasoning puzzles, 50 tasks from the public evaluation set with two attempts per task, as run in Nejumi Leaderboard 4 (Weights & Biases Japan). Accuracy (%), higher is better; the average of ARC-AGI-1 and ARC-AGI-2 is Nejumi 4 - GLP - Abstract Reasoning.
Metric: Accuracy (%). Source: nejumi.ai. Saturation forecast: Around February 2027. 145 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Terra (Max) | 100 | #40 (GPT-5.6 Terra) |
| 2 | GPT-5.6 Sol (Max) | 96 | #22 (GPT-5.6 Sol) |
| 3 | Claude Opus 5 (Max) | 94 | #16 (Claude Opus 5) |
| 4 | GPT-5.4 (2026-03-05) (xHigh) | 84 | #63 (GPT-5.4 (2026-03-05)) |
| 5 | Claude Opus 4.8 (xHigh) | 78 | #39 (Claude Opus 4.8) |
| 6 | Gemini 3.1 Pro (Preview) | 72 | #64 |
| 7 | Claude Opus 4.6 (Thinking) | 72 | #67 (Claude Opus 4.6) |
| 8 | Gemini 3.6 Flash | 70 | #57 |
| 9 | Claude Opus 4.7 (xHigh) | 68 | #51 (Claude Opus 4.7) |
| 10 | GPT-5.4 (2026-03-05) (High) | 68 | #63 (GPT-5.4 (2026-03-05)) |
| 11 | Hy4 preview (High) | 66 | #45 (Hy4 preview) |
| 12 | DeepSeek V4.1 Flash | 64 | #37 |
| 13 | Grok 4.20 0309 (Reasoning) | 64 | #120 (Grok 4.20 0309) |
| 14 | DeepSeek V4 Flash (0731) (Max) | 62 | #84 (DeepSeek V4 Flash (0731)) |
| 15 | Gemini 3.5 Flash | 60 | #66 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=nejumi-4-arc-agi-2 · How It Works · Data refreshed daily, snapshot 2026-11-05.