Benchmarks › ARC-AGI
ARC-AGI
Infer rules for new colored-grid puzzles
95top score · GPT-6 Astra
66models ranked
16open-weights models
Oct 20, 2026last updated
How the test works
1
Task
Infer rules for new colored-grid puzzles
- Kinds of tasksSymbolic interpretation, compositional reasoning, contextual rule application
- In → outColored grids → Answer grid
2
Run
Direct input-output predictor; stateless client
- Token budgetSemi-private evaluation: USD 10,000 total run budget; model-specific token limits
- How runs combineSingle benchmark run; no averaging across runs
3
Score
Every test output grid must be exactly correct, including dimensions
- Better isHigher
- Test setSemi-private tasks sent to model APIs
Sources: arcprize.org, github.com
ARC-AGI-2 · Semi-private leaderboardtop 10 of 66
Full leaderboard with filters →| Model | ARC-AGI-2 · Semi-private | Per task | Price | Speed | Supports | Context |
|---|---|---|---|---|---|---|
95max | $3.26 | $10 / $50 | 69 | 1.05M | ||
94.2max | $0.72 | $2 / $10 | 52 | 1.05M | ||
93.3high | $1.82 | $4 / $20 | 82 | 1M | ||
92.5max | $1.99 | $4 / $20 | 114 | 1.05M | ||
91.3high | $1.99 | $2 / $10 | 84 | 1M | ||
90.4max | $5.86 | $5 / $25 | 55 | 1M | ||
90max | $7.63 | $10 / $50 | 69 | 1M | ||
89.6max | $1.04 | $2 / $10 | 128 | 1.05M | ||
89.2max | $8.75 | $10 / $50 | 67 | 1M | ||
89.2high | $1.24 | $0.75 / $3.75 | 149 | 1M |
Versions and boards on Model Menu
| Board | Models |
|---|---|
ARC-AGI-2 · Semi-privateShown above | 66 |
| 64 | |
| 63 | |
| 61 | |
| 18 |
ARC-AGI questions
What is ARC-AGI?›
Infer rules for new colored-grid puzzles. Created by ARC Prize (original benchmark by François Chollet).
Which model is best on ARC-AGI?›
GPT-6 Astra leads ARC-AGI-2 · Semi-private with 95, ahead of GPT-6.1 Sol (94.2) and Claude Opus 5.5 (93.3).
What is the best open-weights model on ARC-AGI?›
GLM 5.3 Flash with 65.8, then DeepSeek V4 Pro 0813 (61.3).
Which versions of ARC-AGI are there?›
ARC-AGI-2 · Semi-private (66 models), ARC-AGI-1 · Semi-private (64 models), ARC-AGI-2 public evaluation (63 models), ARC-AGI-1 public evaluation (61 models), ARC-AGI-3 · Semi-private (18 models).