The ARC-AGI Performance Rankings

Navigating the Leaderboard

The ARC-AGI benchmark has undergone a significant transformation. It has transitioned from its initial iterations, ARC-AGI-1 and ARC-AGI-2, which focused on ~~passive fluid intelligence~~, to the current ARC-AGI-3. This latest version pushes AI agents to demonstrate real-time adaptation within novel, interactive settings.

The Efficiency Metric

The leaderboard utilizes a scatter plot to highlight the vital correlation between the financial cost of a task and the resulting performance.

"True intelligence isn't just about solving problems, but solving them efficiently with minimal resources."

Mathematically, the efficiency can be viewed as a function of performance relative to cost: $\text{Efficiency} = \frac{\text{Accuracy}}{\text{Cost per Task}}$

Analyzing the Data Categories

The results are segmented into three primary architectural approaches:

System Category	Method	Key Characteristic
Reasoning Systems	Iterative Thinking	Performance vs. Time Trend Lines
Base LLMs	Single-Shot	Raw, unenhanced inference
Kaggle Systems	Optimized Heuristics	Strict budget constraints

1. Reasoning Systems

These entries are visualized as connected trend lines. They track a single model across varying levels of computational "thought." These lines typically exhibit asymptotic behavior, meaning that as the time allocated for reasoning increases, the gains in performance eventually level off.

2. Base LLMs

These represent the "out-of-the-box" capabilities of state-of-the-art models such as GPT-4.5 and Claude 3.7. They utilize standard inference without any extended reasoning loops or external tool-use.

3. Kaggle Systems

These are highly specialized submissions from the Kaggle competition. They are designed for maximum efficiency, adhering to a rigid budget:

Budget: $\le \$ 50$
Scope: 120 evaluation tasks

Governance and Constraints

Verification Policy: Please refer to the official [testing policy] for detailed methodology.

Important Technical Notes:

Cost Ceiling: Any system requiring $>\$ 10,000$ to execute is excluded from the display.
Completion Rule: If a model fails to provide a full output for a test, that task is automatically marked as incorrect.
Status Labels: Results tagged as preview are considered unofficial and may stem from partial testing.
Estimation: Certain ARC-AGI-2 scores are estimates derived from o1-pro pricing and incomplete test sets.

{
  "status": "preview",
  "model": "example-reasoner-v1",
  "cost_per_task": 0.05,
  "accuracy": 0.42
}

Join the Movement: ARC Prize 2026

Ready to contribute to the future of AGI?

Sign up for the official newsletter.
Review the latest contest updates.
Submit your model to the leaderboard.