Benchmarks / Reasoning & Math
ARC-AGI
Abstraction and reasoning over novel grid puzzles; semi-private eval split.
maintained by ARC Prize Foundation · arcprize.org · 1,226 observations on record
Versions
3 versions
| Version | Tasks |
|---|---|
| ARC-AGI-2 current arc-agi-v2 | — |
| ARC-AGI-1 arc-agi-v1 | — |
| ARC-AGI-3 arc-agi-v3 | — |
Current leaderboard
ARC-AGI-2 · Accuracy · top 25 of 157 systems
| # | System | Accuracy % |
|---|---|---|
| 1 | GPT 5.6 Solmax effort OpenAI | 92.5% |
| 2 | GPT 5.6 Solxhigh effort OpenAI | 90.0% |
| 3 | GPT 5.6 Solhigh effort OpenAI | 85.4% |
| 4 | GPT 5.5 (2026-04-22)max effort OpenAI | 85.0% |
| 5 | GPT 5.5 Pro (2026-04-23)high effort OpenAI | 84.6% |
| 6 | Gemini 3 Deep Think 022026 Google DeepMind | 84.6% |
| 7 | GPT 5.5 Pro (2026-04-23)xhigh effort OpenAI | 84.2% |
| 8 | GPT 5.6 Terramax effort OpenAI | 83.9% |
| 9 | GPT 5.4 Pro (2026-03-05)xhigh effort OpenAI | 83.3% |
| 10 | GPT 5.5 (2026-04-22)high effort OpenAI | 83.3% |
| 11 | Gemini 3.1 Pro Preview Google DeepMind | 77.1% |
| 12 | Claude 4.7max effort Anthropic | 75.8% |
| 13 | GPT 5.6 Terraxhigh effort OpenAI | 74.2% |
| 14 | GPT 5.4 (2026-03-05)xhigh effort OpenAI | 74.0% |
| 15 | Gemini 3.5 Flashhigh effort Google DeepMind | 72.1% |
| 16 | Claude Opus 4.8high effort Anthropic | 72.1% |
| 17 | Claude Opus 4.8medium effort Anthropic | 71.7% |
| 18 | GPT 5.5 (2026-04-22)medium effort OpenAI | 70.4% |
| 19 | Claude Opus 4.6high effort · 120K budget Anthropic | 69.2% |
| 20 | Claude Opus 4.6max effort · 120K budget Anthropic | 68.8% |
| 21 | Claude 4.7high effort Anthropic | 68.3% |
| 22 | Claude 4.7medium effort Anthropic | 67.5% |
| 23 | GPT 5.4 (2026-03-05)high effort OpenAI | 67.5% |
| 24 | GPT 5.6 Terrahigh effort OpenAI | 67.1% |
| 25 | GPT 5.6 Solmedium effort OpenAI | 67.1% |
One row per evaluated system — reasoning-effort variants rank separately. Where a system has attr variants for this version (eval splits, style control), only the canonical variant is ranked: split=semi_private, style_control=true, or the unsplit run.