Why this benchmark is useful
ARC Prize is the only benchmark series that repeatedly surfaces new inflection points — from reasoning systems (ARC-AGI-2) to agentic intelligence (ARC-AGI-3).
Abstraction and Reasoning Corpus
ARC-AGI
Fluid, sample-efficient reasoning across the ARC Prize benchmark series — from static grid puzzles (ARC-AGI-1/2) to interactive agentic environments (ARC-AGI-3). Built to resist memorization and reward genuine generalization.
What this does not measure
- Knowledge or language — tasks are deliberately knowledge-free, so scores say nothing about factual accuracy.
- Cost-blind scores mislead — some high ARC-AGI-1 results used enormous per-task compute; always read the $/task footnote.
- Cross-version comparison — ARC-AGI-1, ARC-AGI-2, and ARC-AGI-3 are different tasks with different ceilings. Use the version switcher.
Coverage map
- Task family
- Reasoning
- Format
- Edition-dependent: grid-to-grid puzzles (v1/v2) or interactive turn-based environments (v3). Held-out private sets prevent overfitting.
- Scoring
- % of held-out tasks or environments solved. ARC-AGI-3 measures skill-acquisition efficiency over time.
- Maintainer
- François Chollet / ARC Prize Foundation
How to read the scores
Pick the edition first. ARC-AGI-2 scores are static puzzle accuracy; ARC-AGI-3 scores are environment solve-rates with sparse feedback. Never mix them in a composite.
What it does not cover
- Knowledge or language — tasks are deliberately knowledge-free, so scores say nothing about factual accuracy.
- Cost-blind scores mislead — some high ARC-AGI-1 results used enormous per-task compute; always read the $/task footnote.
- Cross-version comparison — ARC-AGI-1, ARC-AGI-2, and ARC-AGI-3 are different tasks with different ceilings. Use the version switcher.
Scores
4 rowsEvidence ledger
March 2026 interactive benchmark: agents explore novel environments, acquire goals on the fly, and adapt without language instructions. Measures agentic fluid intelligence.
44 rows
0.5%best score
100%human ceiling
99.5 ptsheadroom
4source-checked
1sources
2026-03-20source date
Where the rows land
0255075100
Normalized to this benchmark's axis (0–100). Open the table for raw units.
Newest receipts
- Claude Opus 4.6 (Max)0.5% · ARC-AGI-3 paper (arXiv 2603.24621)
- Gemini 3.1 Pro Preview0.4% · ARC-AGI-3 paper (arXiv 2603.24621)
- GPT 5.4 (High)0.2% · ARC-AGI-3 paper (arXiv 2603.24621)
- Grok 4.20 (Reasoning)0.1% · ARC-AGI-3 paper (arXiv 2603.24621)
Scores use this benchmark's own unit and axis, not a universal quality score.
| # | Model | Release date | Score | Provenance | Trust |
|---|---|---|---|---|---|
| 1 | Claude Opus 4.6 (Max)Anthropic | 2026-03-09 | 0.5% | ARC-AGI-3 paper (arXiv 2603.24621)2026-03-20 | Source-checked |
| 2 | Gemini 3.1 Pro PreviewGoogle | 2026-02-19 | 0.4% | ARC-AGI-3 paper (arXiv 2603.24621)2026-03-20 | Source-checked |
| 3 | GPT 5.4 (High)OpenAI | 2026-04-15 | 0.2% | ARC-AGI-3 paper (arXiv 2603.24621)2026-03-20 | Source-checked |
| 4 | Grok 4.20 (Reasoning)xAI | 2026-03-09 | 0.1% | ARC-AGI-3 paper (arXiv 2603.24621)2026-03-20 | Source-checked |
What it covers
This benchmark sits in the reasoning family. It uses Edition-dependent: grid-to-grid puzzles (v1/v2) or interactive turn-based environments (v3). Held-out private sets prevent overfitting. The score should travel with its task format, scoring method, source date, and benchmark version.
Where it breaks down
Humans solve 100% of environments. Official release rows clustered below 1%. OpenAI's 3 Sep 2026 GPT-6 Astra launch claims 99.9% — vendor-claimed, not an ARC Prize leaderboard row. Do not mark this edition defunct until the official board confirms it.
Tools that report it
Receipts
Sources and further reading
- On the Measure of Intelligence· checked 2026-06-22
- ARC Prize· checked 2026-06-22
- ARC-AGI-3 paper· checked 2026-06-22
- Leaderboard
- ARC-AGI-3
Questions about this benchmark
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
What does ARC-AGI measure?
Fluid, sample-efficient reasoning across the ARC Prize benchmark series — from static grid puzzles (ARC-AGI-1/2) to interactive agentic environments (ARC-AGI-3). Built to resist memorization and reward genuine generalization.
What does a high ARC-AGI score not prove?
A strong ARC-AGI result says nothing about:
- Knowledge or language — tasks are deliberately knowledge-free, so scores say nothing about factual accuracy.
- Cost-blind scores mislead — some high ARC-AGI-1 results used enormous per-task compute; always read the $/task footnote.
- Cross-version comparison — ARC-AGI-1, ARC-AGI-2, and ARC-AGI-3 are different tasks with different ceilings. Use the version switcher.
How is ARC-AGI scored?
% of held-out tasks or environments solved. ARC-AGI-3 measures skill-acquisition efficiency over time. Task format: Edition-dependent: grid-to-grid puzzles (v1/v2) or interactive turn-based environments (v3). Held-out private sets prevent overfitting.
Is ARC-AGI saturated?
ARC-AGI is currently marked Active in the atlas. Ceiling context: Edition-dependent — switch versions above. ARC-AGI-3 remains the official unsolved edition until ARC Prize publishes a confirming leaderboard row.
Can ARC-AGI results be contaminated by training data?
Contamination risk for ARC-AGI is graded Low contamination. Treat every row on this page as a public claim with a source and a date, not as a controlled experiment.
The Pack · Editorial newsletter
New tools in your inbox. Free.
One short email when a tool ships or changes status. No tracking, no third-party analytics. Unsubscribe in one click.