The quest for Artificial General Intelligence (AGI) just got a whole lot more interesting, and challenging! In a groundbreaking development that’s sending ripples through the AI world, the Arc Prize Foundation, spearheaded by AI luminary François Chollet, has unveiled a new, incredibly tough AGI test . Dubbed ARC-AGI-2, this benchmark is designed to truly measure the general intelligence of AI models , and early results are, well, humbling for even the most advanced systems. Why is ARC-AGI-2 the New Hurdle for AI Models? According to a recent blog post from the Arc Prize Foundation, ARC-AGI-2 is proving to be a formidable challenge. Leading “reasoning” AI models like OpenAI’s o1-pro and DeepSeek’s R1 are barely scratching the surface, scoring a mere 1% to 1.3% on the AGI test leaderboard. Even powerhouse non-reasoning models, including GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Flash, are only hitting around the 1% mark. This starkly contrasts with human performance, where panels averaged a solid 60% accuracy on the same ARC-AGI-2 questions. So, what makes this new AI benchmark so different and difficult? Visual Puzzle Problems: ARC-AGI-2 presents AI models with puzzle-like problems involving identifying visual patterns in grids of colored squares. The goal is for the AI to generate the correct “answer” grid based on these patterns. Adaptability Focus: Unlike previous tests, ARC-AGI-2 is specifically designed to force AI models to adapt to entirely new problem types they haven’t encountered during training. This tests genuine problem-solving ability rather than just recall. Efficiency Metric: A key innovation in ARC-AGI-2 is the introduction of an efficiency metric. This means it’s not just about getting the right answer, but also about how efficiently the AI model arrives at the solution. This directly addresses concerns about brute-force computing power masking true intelligence in previous benchmarks. Here’s a comparison of how different types of models are performing: AI Model Type Examples ARC-AGI-2 Score (Approx.) Reasoning AI Models OpenAI o1-pro, DeepSeek R1 1% – 1.3% Non-Reasoning AI Models GPT-4.5, Claude 3.7 Sonnet, Gemini 2.0 Flash Around 1% Human Panels Average participants 60% Source: Arc Prize Leaderboard ARC-AGI-2 vs. ARC-AGI-1: What’s Changed? François Chollet himself stated on X (formerly Twitter) that ARC-AGI-2 is a superior measure of an AI model’s true intelligence compared to its predecessor, ARC-AGI-1. The core aim of the Arc Prize Foundation’s tests remains consistent: to evaluate an AI system’s capacity to learn new skills beyond its training data. However, ARC-AGI-2 tackles a critical flaw in ARC-AGI-1 – the over-reliance on brute-force computation. ARC-AGI-1 , despite being a challenging AI benchmark for years, was eventually “solved” by OpenAI’s o3 (low) model. While o3 (low) achieved an impressive 75.7% on ARC-AGI-1 , demonstrating apparent human-level performance, its score plummeted to a mere 4% on ARC-AGI-2 , even with a significant $200 computing cost per task. This highlights the fundamental shift in focus towards efficiency and genuine intelligence in the new test. Why Does This New AI Benchmark Matter Now? The arrival of ARC-AGI-2 is timely. Many voices within the tech industry, including Hugging Face co-founder Thomas Wolf, have been advocating for more robust and novel benchmarks to truly gauge AI progress, particularly in areas like creativity – a key component of Artificial General Intelligence . The limitations of previous benchmarks were becoming increasingly apparent. Models could achieve high scores through methods that didn’t necessarily reflect genuine intelligence , such as extensive memorization or brute-force computing. ARC-AGI-2 seeks to address these shortcomings head-on, pushing the boundaries of AI evaluation . The Arc Prize 2025: A Challenge to Push AI Intelligence Adding another layer of excitement, the Arc Prize Foundation has announced the Arc Prize 2025 contest. This challenge invites developers to reach an accuracy of 85% on the ARC-AGI-2 test while adhering to a strict cost constraint of just $0.42 per task. This contest is not just about achieving high scores; it’s about fostering efficient and truly intelligent AI models . This new AGI test , ARC-AGI-2 , represents a significant step forward in our ability to measure and understand artificial general intelligence . It’s a stark reminder that while AI models have made incredible strides, true human-level intelligence , characterized by adaptability and efficiency, remains a formidable frontier. The challenge is set, and the race to build truly intelligent AI is on, with ARC-AGI-2 serving as the new, more rigorous yardstick. To learn more about the latest AI benchmark trends, explore our article on key developments shaping AI advancements.
