Skip to content
LLM Atlas

Benchmarks

ARC-AGI-3

activeAgentic tool usecontamination risk: mediumlaunched 2026-04-17

Benchmark metadata from benchwiki — this page mirrors their record so scores can be joined to it. The original is canonical.

What it measures

Interactive ARC environments that score agents by human-relative action efficiency, planning, exploration, and adaptation.

Primary metric
Relative Human Action Efficiency (RHAE) on interactive environments (%)
Judge model
not judged by a model
Human baseline
100The ARC-AGI-3 paper and ARC Prize overview state that environments are 100% human-solvable; the score definition treats 100% as beating every game as efficiently as humans.
Languages
Language-neutral

Health

ARC-AGI-3 scores are not exact-match grid accuracies. They are human-relative efficiency scores over interactive environments. Public-demo, semi-private, and fully-private results should not be mixed without naming the split and budget.

Status evidence
ARC-AGI-3 is newly introduced as an interactive reasoning benchmark. The paper reports that humans solved 100% of environments while frontier AI systems scored below 1% as of March 2026, and the official leaderboard generated on 2026-07-23 still shows low semi-private scores despite visible progress.
Saturated since
not saturated
Refresh cycle
static
Test set public
no

The 25 public demo environments are intentionally easier and do not comprehensively represent the private mechanics. The benchmark uses 55 semi-private and 55 fully-private environments as the main evaluation sets, with the fully private set reserved for tightly guarded competition use.

History & lineage

Hand-maintained, not part of the benchwiki record above — each line cites the paper or leaderboard it comes from.

  1. 2019-11ARC-AGI-1 introduces static colored-grid abstraction tasks.
  2. 2025-03ARC-AGI-2 keeps the static grid format but raises difficulty and improves human calibration and private-set controls.
  3. 2026-04ARC-AGI-3 appears as arXiv:2603.24621v2, introducing interactive turn-based environments and RHAE scoring.
  4. 2026-07The official leaderboard generated on 2026-07-23 lists GPT-5.6 Sol (Max) at 7.78% on the ARC-AGI-3 semi-private axis, with substantially higher cost than static ARC runs.

Further reading

Models scored on this benchmark

The reverse lookup, which is this site’s addition: every model here with a score on this benchmark. Independent measurements and vendor-reported numbers are kept apart rather than interleaved, because they are different claims.

No model on this site has a score on this benchmark yet. That usually means the models that were measured on it are not in the registry, rather than that nobody has run it.

Trajectory

human baselineAnthropic Opus 4.6 (Max) — 0.51 on 2025-12-17 (vendor-reported)Gemini 3.1 Pro (Preview) — 0.42 on 2026-03-05 (vendor-reported)GPT-5.5 (High) — 0.43 on 2026-04-23 (vendor-reported)Claude Opus 4.8 (High) — 1.52 on 2026-06-01 (vendor-reported)GPT-5.6 Sol (Max) — 7.78 on 2026-07-09 (vendor-reported)202520261000
independent · vendor-reported. Points are not joined: the scores come from different models, harnesses and protocol versions, and a line between them would claim a continuity the data does not have.

Where results are tracked

Leaderboards benchwiki lists for this benchmark. These are links out, not sources this site ingests — no score below comes from any of them.

Source

This record is maintained by benchwiki and mirrored here. Read the original. Upstream last updated 2026-07-24; fetched 2026-10-19 by this site.