The ARC Prize Foundation is a non-profit organization that maintains the ARC-AGI benchmark series, administers the annual ARC Prize competitions, and publishes verified evaluations of frontier AI systems against the benchmarks. It describes itself as "a nonprofit advancing open-source artificial general intelligence research through benchmarks & prizes" and states its mission as identifying, measuring and closing "the capability gap between human and artificial intelligence on tasks that are simple for people yet remain difficult for even the most advanced AI systems today" (Source: arcprize.org; arcprize.org).
Origin and organization
The benchmark predates the organization. François Chollet introduced the Abstraction and Reasoning Corpus in the 2019 paper "On the Measure of Intelligence" (ARC-AGI). In June 2024 Chollet and Mike Knoop launched ARC Prize as a $1 million Kaggle competition to open-source a solution to ARC-AGI; the foundation's own account records more than 1,400 teams submitting some 17,000 entries, the $600,000 grand prize going unclaimed, and $75,000 awarded to three research papers and $50,000 to the five highest-scoring teams against the private evaluation set (Source: arcprize.org; arcprize.org).
The competition was constituted as a foundation the following year. The organization's history page records that in 2025 "ARC Prize grew into a non-profit foundation to advance the mission of guiding open-source AGI research," and that the ARC Prize 2025 Kaggle competition, built on ARC-AGI-2, launched with a prize pool of more than $725,000 (Source: arcprize.org). The stated purpose of the shift was to make ARC-AGI "an academic, policy, and industry standard for measuring and guiding AGI" rather than a single annual contest (Source: arcprize.org).
Chollet and Knoop are listed as co-founders and board members, Chollet additionally as creator of ARC-AGI; Greg Kamradt is listed as president and board member. Bryan Landers is listed on the founding team for design and strategy, Lauren Wagner for policy and growth, Hunter Henry for benchmark design, Derek Smith and Ethan Bond for engineering, and Matt Mazur for model testing and analysis. Chollet and Knoop are both associated with the AI research company Ndea; Knoop was previously a co-founder of Zapier (Source: arcprize.org; arcprize.org). The foundation solicits funding through a public donation page and gives team@arcprize.org as its contact (Source: arcprize.org).
Stated position on measuring intelligence
The foundation advances a specific claim about what a benchmark should measure, set out at length on the ARC-AGI pages and summarized in its own about text: that "true AGI requires more than just scaling up existing AI models" and "demands a fundamental shift towards systems capable of genuine fluid intelligence, the ability to adapt to new situations and solve novel problems efficiently" (Source: arcprize.org). Its benchmarks are built around what it calls core knowledge priors, and against the alternative formulation of AGI as the automation of most economically valuable work, which it treats as a goal rather than a measure. The reasoning, the design principle "easy for humans, hard for AI", and the resulting benchmark generations are set out at ARC-AGI.
The organization also states a falsification criterion for its own project: on Kamradt's account, "once we can no longer come up with quantifiable problems that are 'feasible for humans and hard for AI' then we effectively have AGI," and ARC-AGI-2's existence is offered as evidence that this point has not been reached (Source: news.ycombinator.com).
Benchmarks and evaluation role
The foundation publishes the ARC-AGI datasets, holds back semi-private and private evaluation sets, and verifies reported scores. Three generations exist: ARC-AGI-1 (2019), ARC-AGI-2 (2025) and the interactive ARC-AGI-3 (2026), the last of which the foundation published as a technical report (ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence). The ARC-AGI-2 paper (arXiv 2505.11831) is credited to Chollet, Knoop, Kamradt, Landers and Henry Pinkard, and its own reference list describes the foundation as "a nonprofit organization dedicated to fostering open-source scientific progress through enduring AI benchmarks" (Source: arxiv.org).
Beyond the datasets, the foundation functions as a third-party evaluator whose published scores are cited across the model literature. Its December 2024 verification of OpenAI's o3 result on ARC-AGI-1 was run as a live stream co-led by Landers and Kamradt with Chollet and Knoop hosting (OpenAI o-series (o1 → o4-mini); Source: arcprize.org). It has since published evaluations of successive frontier releases, and its results appear on the vault's model and benchmark pages alongside developer-reported figures (AI Benchmarks and Evaluation, Jagged Frontier).
The foundation reports its results with explicit limits on what they establish. Its September 3, 2026 evaluation of GPT-6 Astra states that it is "not claiming that it is AGI," on the ground that the benchmark's environments are deterministic and closed-ended, and separates scores by harness — a Standard harness and a Provider Adapter harness that preserves a model's opaque reasoning state between requests — because the two produce materially different results for the same model (Source: arcprize.org).
Open questions
- Whether the foundation publishes audited financials or discloses its funders is not established from the sources reviewed here; its about page names only a donation page.
- Whether the harness distinction drawn in the Astra evaluation will be applied retrospectively to earlier published scores is not stated.
Relationships
- related: ARC-AGI — the benchmark series the foundation maintains.
- related: François Chollet — co-founder, board member, and creator of ARC-AGI.
- supports: ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence — the foundation's own technical report on the interactive benchmark.
- related: AI Benchmarks and Evaluation — third-party evaluation of frontier models.
- related: Artificial General Intelligence (AGI) — the definitional dispute the foundation's benchmarks are built around.
- related: OpenAI o-series (o1 → o4-mini) — the December 2024 o3 verification.
- related: Jagged Frontier — the uneven-capability pattern ARC-AGI results are cited for.