ARC Prize Foundation

What ARC Prize Foundation does

ARC Prize Foundation is a San Francisco-based nonprofit founded in April 2024 by François Chollet (creator of Keras) and Mike Knoop (Zapier, Ndea) that develops and maintains the ARC-AGI benchmark series — scientifically-grounded evaluations designed to measure AI systems' ability to learn and adapt to novel problems in ways that mirror human fluid intelligence. Its flagship product, ARC-AGI-3 (released March 2026), comprises 135 interactive game environments that test autonomous exploration, hypothesis formation, and plan execution under first-run conditions without instructions or training data, scored via a proprietary Relative Human Action Efficiency (RHAE) methodology that compares AI performance against controlled human baselines established through a 458-participant study. All tested frontier models from Google, OpenAI, Anthropic, and xAI score below 1% on ARC-AGI-3 while humans consistently score 100%.

The foundation distributes its benchmarks and tooling through open-source GitHub repositories, a Python SDK (arc-agi package) capable of 2,000+ FPS local execution, a REST API at three.arcprize.org, community and verified leaderboards, and Kaggle-hosted annual competitions (ARC Prize 2024, 2025, 2026) with over $2 million in prizes. The organization is funded by donations from individuals, foundations, and major AI labs; donors receive no privileged access to evaluation datasets or influence over methodology, with oversight provided by an independent academic panel including professors from NYU, Santa Fe Institute, and Columbia University. The go-to-market is entirely community-led and event-driven, with all benchmark access, API, and datasets offered free of charge; the only disclosed institutional capital is a $500,000 Y Combinator round closed in March 2026.

Name

ARC Prize Foundation

Legal name

ARC Prize, Inc.

Website

https://arcprize.org

Company type

Private

Founded year

2024

Operating status

Operating

Headcount range

1–10 employees

Short description

ARC Prize Foundation is a nonprofit that builds the ARC-AGI benchmark series to measure AI fluid intelligence against human baselines, serving AI labs (Google, OpenAI, Anthropic, xAI) and researchers via open-source tools, Kaggle competitions, and a policy-overseen evaluation platform.

Ownership category

akta.pro rank

Product category

AI Benchmarking Software

NAICS

Grantmaking and Giving Services (81321)

SIC

Services-Testing Laboratories (8734)

akta.pro primary industry

Test Item Banks, Question Authoring & Assessment Design Tools (EDAFADAC)

akta.pro secondary industry

Scholarship & Fellowship Foundations and Trusts (EDADAHAC)

Keywords

• AI benchmarking tools
• AGI research benchmarks
• Abstract reasoning datasets
• AI agent evaluation
• Open-source AI benchmarks

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Operations, Marketing or Sales, Others

Revenue model

1. Donations and Grants: The organization is a nonprofit funded by donations from individuals, foundations, and AI labs. Cash, in-kind donations (e.g., API/compute credits), or other contributions have no influence over testing methodology, roadmap, or publication timing.

Go-to-market motion3 records

Distribution channels4 records

Marketing channels8 records

Core offering

ARC Prize Foundation is a nonprofit organization that builds open-source benchmarks (the ARC-AGI series) designed to measure AI systems' fluid intelligence, skill-acquisition efficiency, and ability to generalize to novel situations. It offers interactive game-based benchmarking environments, a Python SDK and REST API for programmatic evaluation, public/private datasets, leaderboards, and annual Kaggle-hosted prize competitions with over $2M in prizes to drive open-source AGI research.

Product overview

ARC Prize Foundation offers a unified platform of AI benchmarking products centered on the ARC-AGI benchmark series, designed to measure artificial general intelligence through abstract reasoning challenges. The core offering is the ARC-AGI benchmark series (ARC-AGI-1, ARC-AGI-2, and ARC-AGI-3), each providing interactive game-based environments that test AI agents' ability to explore, learn, and adapt to novel situations. The platform includes the open-source ARC-AGI Toolkit (Python SDK) for local and programmatic environment interaction, public/private task datasets, verified and community leaderboards, and the ARC Prize competition hosting over $2M in prizes on Kaggle. The scoring methodology uses Relative Human Action Efficiency (RHAE) to compare AI performance against human baselines.

Differentiator

Problem solved

Functional benefit

Products and services

• ARC-AGI Benchmark Series An open-source benchmark series measuring AI systems' ability to acquire new skills and generalize to novel situations through abstract reasoning challenges. Designed to be easy for humans but hard for AI, identifying key inflection points in AI progress.
• ARC-AGI-3 The third iteration of the ARC-AGI benchmark featuring 135 interactive game environments requiring autonomous exploration, hypothesis formation, and plan execution without instructions. Uses Relative Human Action Efficiency (RHAE) scoring that penalizes brute-force approaches.
• ARC-AGI Toolkit Open-source Python SDK (arc-agi package) for interacting with ARC-AGI-3 environments locally at 2,000+ FPS or via REST API. Supports environment creation, scorecard management, and agent orchestration including multi-agent execution.
• Open Source Datasets Open-source datasets and software for benchmarking AI model capabilities. Includes ARC-AGI-1, ARC-AGI-2, and ARC-AGI-3 datasets along with benchmarking repositories. Public tasks are fully open source; private tasks used for frontier model evaluation.
• ARC-AGI Leaderboards Verified Leaderboard tracks state-of-the-art model performance on ARC-AGI benchmarks, testing only public models from trusted sources with over $10M gross revenue/month. Community Leaderboard allows researchers to share results with light review.
• ARC Prize Competition Annual competition hosted on Kaggle offering over $2M in prizes across 3 tracks for open-source solutions to ARC-AGI-3. Submissions must be open source, runnable via Kaggle notebook in under 12 hours, with runtime costs capped at $10,000.

Quantifiable outcome

• All tested frontier models scored below 1% while humans scored 100%, revealing significant gap between current AI capabilities and general human-level adaptability
• +1 more outcomes

Named customers4 records

Segments3 records

Ideal customer profiles3 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

AI capability6 records

Feature4 records

Partnerships

Four partnerships are on record, tiered core and supporting.

• KagglecoreChannel Partner/ Reseller/ DistributorPartnership with Kaggle to host ARC Prize competitions. Kaggle serves as the platform for the annual ARC Prize competition (ARC Prize 2024, 2025, 2026) with prize payouts administered through the platform. Competitions require one-click runnable notebook submissions.
• Todd Gureckis (Academic Panel)supportingStrategic or Co-development PartnerProfessor of Psychology and Data Science at NYU, serves on the independent academic panel providing external oversight and validation of ARC Prize's benchmarking approach.
• Melanie Mitchell (Academic Panel)supportingStrategic or Co-development PartnerProfessor at the Santa Fe Institute, serves on the independent academic panel providing external oversight of benchmarking methodology, scoring, and publication standards.
• Vishal Misra (Academic Panel)supportingStrategic or Co-development PartnerVice Dean of Computing and AI at Columbia University, serves on the independent academic panel providing external oversight of ARC Prize's benchmarking approach.

Scale indicators5 records

Recent moves6 records

Expansion highlights5 records

Regional players

• Epoch AI: Epoch AI is a research organization tracking AI progress through compute, capability benchmarks, and dataset analysis. It is a regional peer (US-based) providing complementary AI measurement infrastructure, though focused on macro-trend tracking rather than per-model evaluation.

Direct peers

• LMSYS (Chatbot Arena): LMSYS operates the popular Chatbot Arena, a crowdsourced LLM evaluation leaderboard hosted by UC Berkeley researchers. It is a direct peer as a community-driven AI benchmarking organization producing public leaderboards for frontier model comparison.
• MLCommons: MLCommons develops industry-standard AI benchmarks (MLPerf) for ML systems and models. It is a direct peer as a multi-stakeholder benchmark consortium that produces reproducible evaluations for AI capabilities, with similar scientific rigor and industry adoption goals.
• HuggingFace Open LLM Leaderboard: HuggingFace hosts the Open LLM Leaderboard evaluating open-weight models across multiple standardized benchmarks. It is a direct peer as a platform providing reproducible LLM evaluation with public leaderboards, though it focuses on open models rather than closed frontier labs.
• EleutherAI: EleutherAI is a nonprofit AI research lab that develops open-source models and benchmarks (including The Pile). It is a direct peer as an open-science nonprofit focused on reproducible AI evaluation, though with a broader research scope than ARC Prize.

Emerging players

• Scale AI (SEAL Leaderboard): Scale AI operates the SEAL Leaderboard for private model evaluations and expert-driven benchmarking. It is an emerging player that targets enterprise AI evaluation with private, expert-curated test sets — a partially overlapping value proposition with ARC-AGI's private evaluation tier.
• LiveBench: LiveBench is an academic-led benchmark designed to avoid contamination through continuously updated questions. It is an emerging player addressing similar concerns about benchmark saturation that ARC-AGI targets with its RHAE scoring approach.
• SWE-bench (Princeton): SWE-bench is an academic benchmark evaluating AI on real-world software engineering tasks. It is an emerging player focused on a specific capability domain (coding agents), paralleling ARC-AGI's focus on agentic intelligence but in a narrower vertical.

Broad incumbents

• Allen Institute for AI (AI2): AI2 is a nonprofit research institute producing benchmarks (e.g., Tülu, OpenScholar) and AI models. It is a broad incumbent nonprofit in AI research that operates across the same scientific ecosystem, with comparable funding structure (philanthropy + grants) but much larger scale.
• Stanford HELM (Holistic Evaluation of Language Models): Stanford's HELM provides holistic multi-metric evaluation of language models across accuracy, calibration, robustness, and fairness. It is a broad incumbent academic benchmarking initiative with similar transparency goals and academic panel structure.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat5 records

Key risks6 records

Key highlights6 records

Customer concentration

• Xx.com/arcprize
• GitHubgithub.com/arcprize
• YouTubeyoutube.com/channel/UC_rdrp-QkrZn-ce9uCE-0EA

Revenue estimate

Valuation estimate

Number of profiles

Profiles4 records

Funding overview

Funding rounds2 records

Investors1 record

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about ARC Prize Foundation