The ARC Prize 2026’s ARC-AGI-3 competition challenges participants to develop AI agents capable of autonomous interaction, exploration, and decision-making in novel, dynamic environments without instructions, emphasizing real-time reasoning and adaptability. With a total prize pool of $850,000, the competition promotes innovation and transparency through strict rules, open-source requirements, and comprehensive resources, aiming to advance core AI capabilities such as exploration, modeling, goal-setting, and planning.
The ARC Prize 2026 features the ARC-AGI-3 competition, a groundbreaking challenge that tasks participants with developing AI agents capable of interacting with novel, dynamic environments without any provided instructions. This competition marks the first interactive reasoning benchmark, requiring AI agents to explore, adapt, and act efficiently within changing settings. Unlike traditional static tasks, ARC-AGI-3 emphasizes real-time interaction and autonomous decision-making, pushing the boundaries of artificial general intelligence research.
Participants can access comprehensive documentation covering the software development kit (SDK), game environments, and guidelines for building and submitting their AI agents. The competition is hosted on Kaggle, where submissions must be made, and strict rules apply, including no internet access during evaluation and mandatory open sourcing of all code and methods to qualify for prizes. Hardware and compute limitations will be detailed upon the competition’s official launch.
The prize pool for ARC-AGI-3 totals $850,000, with a grand prize of $700,000 awarded to the first eligible agent that achieves a perfect score across the evaluation environments. Additional guaranteed awards include a $75,000 top score prize distributed among the top five performers and $75,000 in milestone prizes for participants who open source their solutions by specified deadlines. These incentives aim to encourage innovation, transparency, and collaboration within the AI research community.
ARC-AGI-3 evaluates four core capabilities essential for advanced AI autonomy: exploration, modeling, goal-setting, and planning and execution. Exploration requires agents to actively gather information by interacting with their environment. Modeling involves creating generalizable world models from raw observations. Goal-setting challenges agents to independently identify desirable future states without explicit instructions, while planning and execution demand strategic action sequencing and adaptability to environmental feedback.
The competition builds on insights from previous ARC-AGI challenges, including the ARC-AGI-3 Developer Preview, which introduced the interactive reasoning format. Interested participants and followers can stay informed through various channels such as newsletters, Discord, Twitter, YouTube, and GitHub. The ARC Prize organization provides extensive resources, including leaderboards, research materials, and community engagement opportunities, fostering a vibrant ecosystem for advancing artificial general intelligence.