A benchmark of novel puzzles that the ARC Prize Foundation keeps as a test of general intelligence. Humans score 100 percent on it, and row 7 of the Scorecard reads the best model’s score. Chapter 3.
From The AI Compute Race. Read the book
Where the book uses it
- Chapter 3 — The Missing Researcher
The dated promise of an AI that does its own research, what the laboratories have published, and what the public record can and cannot yet check.
Where the record uses it
- Scorecard indicator ARC-AGI-3
Sources that use it
- ARC Prize Foundation, “Launching ARC-AGI-3,” 25 March 2026 (“Humans score 100%. Frontier AI scores 0.51%”). In the Sources
- ARC Prize Foundation, “Verified Leaderboard” for ARC-AGI-2 and ARC-AGI-3, read 29 August and 4 September 2026 (on 29 August: Claude Opus 5 (High) 30.16% on ARC-AGI-3, GPT-5.6 Sol (Max) 92.50% at $1.44 a task on ARC-AGI-2, Human Panel 100% at $17.00; on 4 September, on ARC-AGI-2: GPT-6 Astra (Max) 95.00% at $1.12 a task; the title is the board’s own heading on the site, and the figures were read from the site’s own data files, which the client-rendered page draws its figures from: the live v2.json and v3.json, generated 2026-09-04T14:38Z, and the archived copies named next, in the order of the URLs below after the board itself, which is the first URL — for the board as it stood on 29 August 2026, the Wayback Machine’s snapshot of 28 August 2026 of v3.json, built 2026-08-21, and for ARC-AGI-2 the snapshot of 19 August 2026 of v2.json, built 2026-08-13, which the later file extends without removing a row; and the live v3.json as generated 2026-10-05T19:34Z, 87 evaluations: GPT-6 Astra (Max) 62.71% on the Standard harness, GPT-6 Astra - Provider Adapter (High) 99.95% and the same adapter at Max 98.55%, which the chapter prints as 98.6, and above 96% at each of its six settings). In the Sources
- BenchLM, ARC-AGI-3 leaderboard snapshot, 28 August 2026 (Claude Opus 5 at 30.2 percent). In the Sources
- Greg Kamradt (ARC Prize Foundation), “OpenAI’s GPT-6 Astra on ARC-AGI-3,” ARC Prize blog, 3 September 2026 (“GPT-6 Astra scores 62.7% for $26K on ARC-AGI-3 Semi-Private with our Standard harness, and 99.9% for $19K with a Provider Adapter harness”; the post’s own table gives those two runs as 62.7% at $26,098 at max reasoning effort on the Standard harness and 99.9% at $18,817 at high effort on the Provider Adapter harness; “used fewer actions than the human baseline on 96.0% of levels”; “its ability to turn unfamiliar environments into compact symbolic world models”; “we are not claiming that it is AGI”). In the Sources
- Greg Kamradt, “Measuring Human Performance on ARC-AGI-3,” ARC Prize blog, 14 April 2026 (the human baseline: “ARC-AGI-3 is 100% solvable by humans”). In the Sources
Related terms
From the glossary of The AI Compute Race.
How to cite this page
Dryburgh, Lee S. “ARC-AGI-3.” In The AI Compute Race. DRYBURGH. https://dryburgh.com/vol1/glossary/arc-agi-3 (accessed 2026-11-01).
BibTeX
@misc{dryburgh-arc-agi-3-2026,
author = {Dryburgh, Lee S.},
title = {{ARC-AGI-3}},
howpublished = {In \emph{The AI Compute Race}, DRYBURGH},
year = {2026},
url = {https://dryburgh.com/vol1/glossary/arc-agi-3},
note = {Accessed 2026-11-01}
}