Abstract reasoning · %
ARC-AGI-2
Harder second version of ARC-AGI (2025), designed so that brute-force search is not enough.
Active Contamination risk: Low 10 results · 4 models Updated Oct 22, 2026 Newest result: 2026-10-22 Official leaderboard ↗
What it measures
% solved on the semi-private set, with cost per task.
Limitations
Only the top entries are cited (terms). Systems with large test-time compute can buy accuracy with money.
Contamination risk: Low
Semi-private set.
Leaderboard
4 models
| # | Model | Provider | Score | Rank | Setting | Source type | Date |
|---|---|---|---|---|---|---|---|
| 1 | GPT-6 Astra | OpenAI | 95.0% | 1 | Semi-private set; $1.12/task; CoT“GPT-6 Astra (Max)” | Official leaderboard | Oct 22, 2026 |
| 2 | GPT-6.1 Sol | OpenAI | 94.2% | 2 | Semi-private set; $0.25/task; CoT“GPT-6.1 Sol (Max)” | Official leaderboard | Oct 22, 2026 |
| 3 | Claude Opus 5.5 | Anthropic | 93.3% | 4 | Semi-private set; $0.41/task; CoT“Claude Opus 5.5 (High)” | Official leaderboard | Oct 22, 2026 |
| 4 | GPT-5.6 Sol | OpenAI | 92.5% | 5 | Semi-private set; $1.44/task; CoT“GPT-5.6 Sol (Max)” | Official leaderboard | Oct 22, 2026 |
Data sources & licences
| Source | Licence / terms | How we reuse it | Last fetched |
|---|---|---|---|
| ARC Prize leaderboard ARC Prize Foundation, arcprize.org/leaderboard | ARC Prize Terms: no copying/aggregation for commercial purposes without permission | Restricted – only top entries cited top 10 per set only (terms) | Oct 22, 2026 |
FAQ
What does ARC-AGI-2 measure?
Harder second version of ARC-AGI (2025), designed so that brute-force search is not enough. % solved on the semi-private set, with cost per task.
How reliable is ARC-AGI-2?
Only the top entries are cited (terms). Systems with large test-time compute can buy accuracy with money. Contamination risk: Low – Semi-private set.