ARC Prize Foundation
Nonprofit that builds AGI benchmarks (ARC-AGI) frontier labs use to grade their own models.
On the record
Checkpoint: ARC Prize 2026 results across ARC-AGI-2, ARC-AGI-3, and Paper Prize tracks
The call
A four-person team keeps building the tests that grade AI progress, racing to redesign each one before the models it measures outgrow it.
ARC Prize Foundation creates tests to measure how smart AI systems are getting, and other organizations can't fairly grade their own AI because they want to sell it. This creates real demand for ARC's work. But here's the problem: AI gets better so fast that tests become too easy in about a year or two, so ARC has to constantly build new ones. With only four people, ARC is in a race against time, against other testing groups like METR, and against the possibility that big AI labs will just get together and create their own standard instead. The biggest thing to watch is whether ARC can become the official standard that everyone has to use, which would lock in their importance before someone else claims that spot.
Why
01
AI pressure: Frontier models themselves (e.g.. GPT-5.2 reportedly jumped from ~54% to ~98% on ARC-AGI-2 in four months, saturating the exact test it sells. Could AI do the core job? Only partly.
02
Bigger players: METR, Epoch AI, and lab-backed evaluation consortiums. Peers are racing for the same role as the reference AI standard, including inside NIST's CAISI program. Could a giant copy it? No, they are blocked.
03
What protects it. Frontier labs cannot credibly grade their own AGI claims, so an outside scorer stays needed even as each test expires.
What would change this call
Whether NIST CAISI names an official standard evaluator within 12-18 months. Shows if ARC Prize locks in the reference-metric role or loses it to METR, Epoch AI, or a lab consortium.
Time between ARC-AGI-3 saturation and ARC-AGI-4 shipping. Shows whether the 4-person team can outrun model progress or falls behind the redesign treadmill.
Any public dispute over donor lab influence on ARC Prize scoring or design. Would test whether the foundation's independence claim holds under scrutiny.
Time between ARC-AGI-3 saturation and ARC-AGI-4 shipping. Shows whether the 4-person team can outrun model progress or falls behind the redesign treadmill.
Any public dispute over donor lab influence on ARC Prize scoring or design. Would test whether the foundation's independence claim holds under scrutiny.
Key numbers
- ARC-AGI-1 stood unbeaten about 5 years before falling
- ARC Prize 2026 prize pool: $2,000,000 across three tracks
- Top verified model scored 37.6% on ARC-AGI-2
Named risks
- Benchmark saturation: ARC-AGI-1 solved after ~5 years, ARC-AGI-2 solved within ~1 year, forcing constant redesign.
- Donor independence perception risk from accepting funding tied to AI labs it evaluates.
- Only 4-person team must outrun model progress and ship new benchmarks (ARC-AGI-3, future ARC-AGI-4).
- Mindshare competition from METR, Epoch AI, LMArena for the reference-metric role.
- Dependency on shifting US government AI-standards priorities (NIST CAISI relationship).
Rivals, as read in July 2026
| Rival | Direction | Evidence |
|---|---|---|
| METR | contesting | Spun from Alignment Research Center Dec 2023, competes for AI capability/safety evaluation mindshare |
| Epoch AI | contesting | Independent nonprofit running rigorous AI capability benchmarks like FrontierMath |
| LMArena | contesting | Crowdsourced human-preference leaderboard offers a rival model-ranking signal |
| Frontier labs (OpenAI, Anthropic, Google, xAI) | gaining on | Labs increasingly publish own internal evals, reducing reliance on external benchmarks like ARC's |
Questions people ask
- Is ARC Prize Foundation profitable?
- ARC Prize Foundation is a nonprofit, not a profit-seeking company. It runs on donations from individuals, foundations, and AI labs, stating on Jan 8, 2025 that before becoming a formal foundation it was funded exclusively by co-founders Francois Chollet and Mike Knoop.
- Who founded ARC Prize Foundation?
- ARC Prize was co-founded by Francois Chollet, creator of the original 2019 ARC-AGI benchmark and Keras, and Mike Knoop, Zapier co-founder, in April 2024. It became a formal nonprofit in early 2025 with Greg Kamradt as President, per the Jan 8, 2025 announcement.
- Could AI progress make ARC Prize's benchmarks obsolete?
- Yes, this is the core risk. ARC-AGI-1 stood unbeaten for about 5 years until OpenAI's o3 tested against it live in December 2024. ARC-AGI-2 launched March 2025 specifically to resist the same saturation, forcing the foundation into continuous benchmark redesign.
- Who are ARC Prize Foundation's biggest competitors?
- Peer independent AI evaluators include METR (spun from the Alignment Research Center in December 2023), Epoch AI, and LMArena, all competing for the role of trusted reference benchmark. Frontier labs like OpenAI and Google also increasingly publish their own internal evals.
- Does the US government use ARC Prize's benchmarks?
- ARC Prize is listed as a provider of AI benchmarks to NIST's Center for AI Standards and Innovation (CAISI), per arcprize.org. It also submitted input to NIST's AI Standards 'Zero Drafts' pilot on April 15, 2025, and NIST announced a related AI Agent Standards Initiative on Feb 17, 2026.
- How big is ARC Prize's 2026 competition prize pool?
- ARC Prize 2026 has a $2,000,000 total prize pool split across the ARC-AGI-2 track, the new ARC-AGI-3 track, and a Paper Prize, according to arcprize.org's 2026 competition page. This is up from over $725,000 in the 2025 competition.
Timeline
- 2026-03-26TechCrunch reports OpenAI, Anthropic, Google already use ARC Prize benchmarks, at YC W26 Demo Day
- 2026-02-17NIST announces AI Agent Standards Initiative
- 2025-07-29ARC Prize issues statement on US AI Action Plan, warns of China benchmark standards push
- 2025-07-18ARC-AGI-3 preview agent competition begins, runs through Aug 19, 2025
- 2025-04-15ARC Prize submits input to NIST AI Standards Zero Drafts pilot
- 2025-03-12ARC Prize submits benchmark recommendations to US OSTP AI Action Plan
- 2025-03ARC-AGI-2 benchmark launched, designed to resist brute-force memorization
- 2025-01-08ARC Prize Foundation nonprofit formally announced with Greg Kamradt as President
- 2024-12-20OpenAI's o3 tested live against ARC-AGI-1, breaking 5-year unbeaten streak
- 2024-04ARC Prize co-founded by Francois Chollet and Mike Knoop
Revision history
27 July 2026
First assessed. Defensible for now. The text above is fixed as of this date and is not edited afterwards.
In the same band
Sources
- ARC Prize Foundation - a North Star for AGI (Jan 8, 2025)
- ARC Prize Foundation
- ARC Prize
- 16 of the most interesting startups from YC W26 Demo Day (March 26, 2026)
- ARC Prize 2024: Technical Report
- ARC Prize 2025 Results and Analysis
- ARC Prize 2026
- GPT-5.2 & ARC-AGI-2: A Benchmark Analysis of AI Reasoning (Dec 12, 2025)
- ARC Prize Blog
- A New Challenge for Frontier AI Reasoning Systems
- ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence
- Benchmarks Matter: ARC Prize Foundation's Recommendations for the U.S. AI Action Plan (March 12, 2025)