ARC-AGI-3
Interactive ARC environments that score agents by human-relative action efficiency, planning, exploration, and adaptation.
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.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| Anthropic Opus 4.6 (Max) | 0.51% | 2024-06-01 | vendor-reported | Source ↗ |
| Gemini 3.1 Pro (Preview) | 0.42% | 2024-06-01 | vendor-reported | Source ↗ |
| GPT-5.5 (High) | 0.43% | 2024-06-01 | vendor-reported | Source ↗ |
| Claude Opus 4.8 (High) | 1.52% | 2024-06-01 | vendor-reported | Source ↗ |
| GPT-5.6 Sol (Max) | 7.78% | 2024-06-01 | vendor-reported | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Human trial-and-error discovery and completion rate on multi-step dynamic grid environments.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.Relative Human Action Efficiency (RHAE) on interactive environments (%)Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.$5 – $20 USD for full benchmark evaluation run on frontier APIs.
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.lm_eval --model hf --model_args pretrained=<model_path> --tasks arc-agi-3 --batch_size autoopencompass --datasets arc-agi-3 --models <model_config># Standard API Evaluation Loop
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-4o",
messages=[{"role": "user", "content": prompt}],
temperature=0.0,
)When publishing results for ARC-AGI-3, always report the exact prompt template, few-shot exemplar ordering, sampling temperature (temperature=0), maximum reasoning budget tokens, and the precise timestamped model snapshot ID.
Contamination & Memorization Analysis
Audit of pretraining exposure risks, memorization vectors, and refresh policies.Static fixed snapshot
Public on web / HuggingFace