The North Star for AGI

Level Unlocked: Over $2.4M Raised!

Additional contributions help us expand our work and create more impactful benchmarks!

Donations are tax-deductible.

Thank You To Our Donors

ARC Prize Foundation

The ARC Prize Foundation is a nonprofit organization dedicated to open scientific progress through enduring AI benchmarks.

Other benchmarks create increasingly complex challenges, but they mask a fundamental truth: AI systems still lack extremely basic capabilities that humans possess.

ARC-AGI is unique among all benchmarks in the world by targeting a critical gap, problems humans can solve but AI systems cannot.

Effective benchmarks don't show gradual improvement, they are capability assertions illuminating step function changes. Our first benchmark, ARC-AGI-1, proved this: it remained unsolved for 5 years until a breakthrough finally occurred.

But we have more work to do.

ARC-AGI-2 Preview

ARC-AGI-2 has been in the works for years. We recently completed both our human calibration testing and frontier AI system testing. Early results show humans achieve a 100% solve rate while still being extremely difficult for state-of-the-art AI systems. We are seeing ARC-AGI-2 computer solve rates between 10-20% even with expensive frontier reasoning systems like o3. This gap demonstrates these systems do not yet have human-level general intelligence.

The final step before publishing (March 2025) is to make final task bucket selections and write the technical report. We're execited to share ARC-AGI-2 alongside the launch of ARC Prize 2025.

ARC-AGI-1 was designed to be challenge for deep-learning based AI systems. ARC-AGI-2 is designed to be challenging for "reasoning" systems -- while maintaining a 100% pass solve for humans. We've studied frontier reasoning systems over the past several months and found several interesting task characteristics amongst unsolved tasks:

• Symbolic Interpretation: We found that frontier systems struggle with tasks requiring symbols to be interpreted as having meaning beyond their visual patterns. Systems attempted symmetry checking, mirroring, and other transforms, and even recognized connecting element, but failed to assign semantic significance to the symbols themselves.
• Multi-step Composed Rules: Frontier systems struggle with tasks requiring sequential application of different rules, especially when the pattern isn't immediately obvious.
• Contextual Rule Application: Systems struggle with tasks where rules must be applied differently based on context. They often fixate on superficial patterns rather than understanding the underlying selection principles that humans easily grasp.

Look out for the full ARC-AGI-2 dataset, testing methodology, and design details in our technical report coming next month.

ARC-AGI-3 Design

Our expectation is that ARC-AGI-2 will not endure 5 years like the original benchmark and we've begun design of the third version of the benchmark, intended to launch in 2026. Our design goal for ARC-AGI-3 is to endure 3+ years.

The primary use of donations will be used to fund development of ARC-AGI-3.

We're exploring ARC-like games which codifies a formal measure of human-comparable efficiency directly into the benchmark design, requires new capabilities like goal orientation, and raises the bar for on-the-fly adaptation to resist brute-force data augmentation. All while maintaining the classic ARC feel.

And of course, the games must be fun!

Note: we're still exploring and these details may change.

Beyond

Our work goes beyond creating benchmarks. We deploy them: measure AI's real capabilities, enable open progress through a global competition, and connect top experts with the resources to build systems that empower human progress.

We are uniquely positioned to provide definitive and independent tests that measure when AI can truly generalize like a human, but we need your help. Partner with us to shape the future.

Donations are tax-deductible.

Benchmarks

As long as there are tasks which humans can do, but AI cannot, we do not have AGI. We measure and quantify that gap.

ARC-AGI-1: Published in 2019. Endured for over 5 years. It was the basis for 4 global AI competitions (over 2K teams submitted solutions). Highlighted significant progress in the field.

ARC-AGI-2: Arriving March 2025. Similar format to ARC-AGI-1, but with more signal towards general intelligence. This includes tasks with more symbolic interpretation, local interactions, framing generalizations, and multi-step reasoning. These tasks are still feasible for humans, but are still challenging (10-20%) for state-of-the-art systems. ARC-AGI-2 will be the basis for the ARC Prize 2025 competition.

ARC-AGI-3: Coming 2026. A new paradigm to measure AI systems that can reason. Let the games begin.