ARC-AGI-3

2026-04-09 · Source: ARC Prize · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, long

Summary

ARC Prize has released ARC AGI 3, a new benchmark designed to measure progress towards Artificial General Intelligence (AGI) by focusing on human-like learning capabilities. Unlike previous versions that used static puzzles, ARC AGI 3 places AI agents in interactive, novel game environments without instructions, forcing them to explore, acquire goals, build world models, and learn continuously. Each of the hundreds of handcrafted games and nearly a thousand levels is human-solvable, with human efficiency baselines established by general public testers. Current frontier AI models score under 1% on ARC AGI 3, exhibiting limitations in anticipating future events and learning from past experiences. The ARC Foundation aims to close this human-AI gap, defining AGI as the point where AI can solve all problems humans can, and has announced a $2 million ARC Prize 2026 competition on Kaggle for both ARC AGI 3 and the final year of ARC AGI 2.

Key takeaway

For research scientists focused on advancing AI towards AGI, you should prioritize developing agents capable of continuous learning and dynamic world model construction. ARC AGI 3 provides a critical, unsaturated benchmark that isolates current AI limitations in exploration, goal acquisition, and adaptive learning, offering a clear target for future research and development. Engaging with the ARC Prize 2026 competition on Kaggle could accelerate progress and validate novel approaches.

Key insights

General intelligence is defined by the ability to learn new skills, not just perform existing ones.

Principles

• Benchmarks should target problems humans solve, AI cannot.
• AGI implies a finite set of human-solvable problems for AI.
• Learning efficiency is a key metric for general intelligence.

Method

ARC AGI 3 uses novel, instruction-less interactive game environments to test AI's ability to explore, acquire goals, build world models, and learn continuously, benchmarking against human efficiency.

In practice

• Explore ARC AGI 3 public games on arcprize.org.
• Participate in the $2 million ARC Prize 2026 competition.
• Develop agents for interactive, continuous learning environments.

Topics

• General Intelligence
• ARC AGI Benchmarks
• Human-AI Gap
• ARC AGI 3
• Algorithmic Learning Efficiency

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• [AINews] Humanity's Last Gasp — AI's progress in automation paradoxically correlates with increased human workload, highlighting a gap in generalizable learning.
• Francois Chollet + Mike Knoop | ARC Prize @ MIT — ARC benchmarks measure "micro-AGI" properties like efficient interactive learning and goal discovery in novel, small-scale environments.
• Why AI Needs Better Benchmarks — AI benchmarks must continuously evolve to measure true general intelligence beyond memorization and narrow task performance.
• AGI: Francois Chollet + Sam Altman — AI's rapid advancement necessitates evolving social contracts and re-evaluating human agency and intelligence.
• Live Q&A With the Hosts: Measuring Progress Toward AGI - Cognitive Abilities Hackathon — Robust AGI evaluation requires targeted benchmarks for specific cognitive faculties, moving beyond general problem-solving.
• Rumors of AGI’s arrival have been greatly exaggerated — Benchmark performance in AI is not sufficient evidence for Artificial General Intelligence (AGI).

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ARC Prize.