ARC-AGI-1
The original ARC-AGI-1 abstract-reasoning puzzle benchmark (semi-private set): few-shot grid transformations that are easy for humans but resist memorization. Largely cleared by 2026 frontier reasoning models, which is what motivated the harder ARC-AGI-2.
What this benchmark measures
The original ARC-AGI-1 abstract-reasoning puzzle benchmark (semi-private set): few-shot grid transformations that are easy for humans but resist memorization. Largely cleared by 2026 frontier reasoning models, which is what motivated the harder ARC-AGI-2.
Rows on this page are sourced from public benchmark artifacts, leaderboard exports, or source-linked model reports. Each row keeps benchmark version, source model name, and available run details attached to the score.
The metric shown here is accuracy. It should be interpreted within ARC-AGI-1, not compared as part of a site-wide ranking.
What to be careful about
Largely saturated at the frontier; public and semi-private splits differ, so keep the reported effort/compute as run context.
Frequently asked
What is ARC-AGI-1?
The original ARC-AGI-1 abstract-reasoning puzzle benchmark (semi-private set): few-shot grid transformations that are easy for humans but resist memorization. Largely cleared by 2026 frontier reasoning models, which is what motivated the harder ARC-AGI-2. It is a reasoning benchmark measured by accuracy.
What does accuracy mean on ARC-AGI-1?
ARC-AGI-1 reports accuracy (%); higher is better. Scores are shown only within ARC-AGI-1 and are never averaged with other benchmarks.
What is the top reported ARC-AGI-1 score?
Gemini 3.1 Pro Preview has the top reported score on ARC-AGI-1: 98% (accuracy).
Why do ARC-AGI-1 scores differ across runs?
Harness, scaffold, reasoning effort, and prompt setup change results, so two runs of the same model can differ. evals.report keeps each score with its run context so the differences stay visible.
Does evals.report rank models across benchmarks?
No. ARC-AGI-1 scores are shown within their own metric; evals.report never combines benchmarks into a composite ranking or a single "best model".