ReasoningNearing Saturation
Trust Score:
75
ARC-AGI-2
The harder static ARC successor: human-calibrated colored-grid abstraction tasks, now pressured by 2026 frontier systems.
Launched: —Refresh: static
Status Assessment (nearing-saturation):
ARC-AGI-2 remains the current static colored-grid ARC benchmark, but the official ARC Prize leaderboard generated on 2026-07-23 lists GPT-5.6 Sol Max at 92.5% and GPT-5.6 Sol xHigh at 90.0% on the ARC-AGI-2 semi-private axis. Because competition/private-set validation and efficiency constraints still matter, this page marks it nearing saturation rather than fully saturated.
Performance Timeline
Longitudinal progression of model scores against human baselines.Performance & Historical Trajectory
Empirical score progression across model release dates and evaluation rounds.
Independent Vendor-Reported Human Baseline (82.0%)
| Model | Score | Date | Source Type | Provenance |
|---|---|---|---|---|
| DeepSeek V4-Flash | 67.5% | 2026-09-12 | independent | Source ↗ |
| DeepSeek V4-Pro | 71.8% | 2026-09-12 | independent | Source ↗ |
| Claude Fable 5.1 (Hybrid Reasoning) | 70% | 2026-09-12 | independent | Source ↗ |
| Gemini 3.7 Flash | 72.5% | 2026-09-10 | independent | Source ↗ |
| Grok 4.6 | 72.9% | 2026-09-09 | independent | Source ↗ |
| GPT-5 | 73.9% | 2026-09-04 | independent | Source ↗ |
| DeepSeek V4.1 Flash | 69% | 2026-08-29 | independent | Source ↗ |
| GPT-6 Astra (max) | 71.5% | 2026-08-25 | independent | Source ↗ |
| Kimi K3 | 73.2% | 2026-08-23 | independent | Source ↗ |
| Claude Opus 5 | 74.5% | 2026-08-21 | independent | Source ↗ |
| Claude Sonnet 5 | 72.5% | 2026-08-21 | independent | Source ↗ |
| Gemini 3.5 Flash-Lite | 66.7% | 2026-08-18 | independent | Source ↗ |
| Gemini 3.6 Flash | 71.4% | 2026-08-18 | independent | Source ↗ |
| GPT-5.6 Sol | 74.9% | 2026-08-06 | independent | Source ↗ |
| GPT-5.5 Thinking | 73.7% | 2026-08-06 | independent | Source ↗ |
| MiniMax M3 | 68.6% | 2026-07-23 | independent | Source ↗ |
| QwQ Plus | 71% | 2026-07-16 | independent | Source ↗ |
| Qwen3-235B | 71.4% | 2026-07-16 | independent | Source ↗ |
| Grok 4 | 73.7% | 2026-07-13 | independent | Source ↗ |
| GLM-5.2 | 69.4% | 2026-07-10 | independent | Source ↗ |
| Claude Fable 5 | 75.3% | 2026-07-07 | independent | Source ↗ |
| Gemini 3 Pro | 71.8% | 2026-06-15 | independent | Source ↗ |
| o3 | 72.9% | 2026-01-17 | independent | Source ↗ |
| Mistral Large 3 | 66.3% | 2025-12-10 | independent | Source ↗ |
| Gemma 3 27B Preview | 65.5% | 2025-05-10 | independent | Source ↗ |
| Llama 4 Maverick | 70.2% | 2025-05-03 | independent | Source ↗ |
| Llama 4 Scout | 69% | 2025-05-03 | independent | Source ↗ |
| Gemini 2.5 Pro | 68.6% | 2025-04-22 | independent | Source ↗ |
| QwQ 32B | 67.2% | 2025-04-02 | independent | Source ↗ |
| GPT-4.5 | 66.3% | 2025-03-27 | independent | Source ↗ |
| Claude 3.7 Sonnet | 69.5% | 2025-03-24 | independent | Source ↗ |
| o3-preview-low (CoT + search/synthesis) | 4% | 2025-03-24 | vendor-reported | Source ↗ |
| ARChitects (Kaggle 2024 winner) | 3% | 2025-03-24 | independent | Source ↗ |
| r1 / r1-zero (single CoT) | 0.3% | 2025-03-24 | vendor-reported | Source ↗ |
| GPT-4.5 (pure LLM) | 0% | 2025-03-24 | vendor-reported | Source ↗ |
| Grok 3 | 71.8% | 2025-03-17 | independent | Source ↗ |
| Gemini 2.5 Flash (Thinking) | 66.5% | 2025-03-10 | independent | Source ↗ |
| Gemini 2.0 Flash | 64% | 2025-03-05 | independent | Source ↗ |
| Sonar Reasoning Pro | 70.2% | 2025-03-05 | independent | Source ↗ |
| Gemini 2.0 Pro | 66.7% | 2025-03-05 | independent | Source ↗ |
| o3-mini | 68.8% | 2025-02-28 | independent | Source ↗ |
| Qwen 2.5 Max | 65.9% | 2025-02-25 | independent | Source ↗ |
| Ollama DeepSeek-R1 (Q4_K_M) | 71.4% | 2025-02-19 | independent | Source ↗ |
| Kimi k1.5 | 66.3% | 2025-02-17 | independent | Source ↗ |
| Gemini 2.0 Flash Thinking | 67.9% | 2025-02-17 | independent | Source ↗ |
| DeepSeek-R1 | 68.6% | 2025-02-17 | independent | Source ↗ |
| MiniMax-Text-01 | 63.2% | 2025-02-12 | independent | Source ↗ |
| Codestral 25.01 | 57.7% | 2025-02-11 | independent | Source ↗ |
| DeepSeek-V3 | 63.9% | 2025-01-23 | independent | Source ↗ |
| Phi-4 | 59.3% | 2025-01-09 | independent | Source ↗ |
| Sonar Pro | 65.1% | 2025-01-07 | independent | Source ↗ |
| Ollama Llama 3.3 (Q4_K_M) | 64.7% | 2025-01-05 | independent | Source ↗ |
| Llama 3.3 70B | 62% | 2025-01-03 | independent | Source ↗ |
| Llama 3.3 70B Instruct | 62.8% | 2025-01-03 | independent | Source ↗ |
| o1 | 67.7% | 2025-01-02 | independent | Source ↗ |
| Amazon Nova Pro | 63.2% | 2024-12-31 | independent | Source ↗ |
| Amazon Nova Lite | 57.7% | 2024-12-31 | independent | Source ↗ |
| QwQ-32B Preview | 67.1% | 2024-12-26 | independent | Source ↗ |
| GPT-4o | 63.6% | 2024-12-18 | independent | Source ↗ |
| Sonar | 61.6% | 2024-12-13 | independent | Source ↗ |
| Qwen 2.5 Coder 32B | 60.8% | 2024-12-10 | independent | Source ↗ |
| Hunyuan-Large | 64.7% | 2024-12-03 | independent | Source ↗ |
| Claude 3.5 Sonnet (v2) | 65.1% | 2024-11-19 | independent | Source ↗ |
| Claude 3.5 Sonnet | 66.3% | 2024-11-19 | independent | Source ↗ |
| Claude 3.5 Haiku | 58.5% | 2024-11-19 | independent | Source ↗ |
| Yi-Lightning | 64% | 2024-11-12 | independent | Source ↗ |
| GLM-4-Plus | 62.4% | 2024-09-17 | independent | Source ↗ |
| Gemma 2 2B | 43.7% | 2024-08-28 | independent | Source ↗ |
| Mistral Large 2 | 62.2% | 2024-08-21 | independent | Source ↗ |
| Mistral Large 2 | 60.1% | 2024-08-21 | independent | Source ↗ |
| Llama 3.1 405B | 63.2% | 2024-08-20 | independent | Source ↗ |
| GPT-4o mini | 54.6% | 2024-08-15 | independent | Source ↗ |
| ERNIE 4.0 Turbo | 62.4% | 2024-07-26 | independent | Source ↗ |
| Gemma 2 9B | 55.4% | 2024-07-25 | independent | Source ↗ |
| Gemma 2 27B | 60.1% | 2024-07-25 | independent | Source ↗ |
| GPT-5.6 Sol (Max) | 92.5% | 2024-06-01 | vendor-reported | Source ↗ |
| GPT-5.5 (xHigh) | 85% | 2024-06-01 | vendor-reported | Source ↗ |
| Claude 4.7 (Max) | 75.8% | 2024-06-01 | vendor-reported | Source ↗ |
| NVARC | 24% | 2024-06-01 | independent | Source ↗ |
| Snowflake Arctic | 58.5% | 2024-05-22 | independent | Source ↗ |
| DBRX Instruct | 59.3% | 2024-04-24 | independent | Source ↗ |
Human Baseline & Difficulty Horizon
Calibrated human reference points, specialist benchmarks, and ceiling thresholds.Measured Human Score82.0%Domain Expert Baseline
Baseline Protocol & Interpretation
Human solve rate on updated 2025 ARC-AGI visual logic puzzles with increased transformation complexity.
Metric & Scoring Methodology
Verification protocols, aggregation formulas, and specialized metric variants.Primary Metric:
task success rate on the public evaluation split (%)Scoring Engine:exact-match
Dataset & Compute Cost
Evaluation volume, public availability, API pricing, and local hardware requirements.Total Dataset Size120Annotated evaluation items
Public Test SetPublicOpenly mirrored on repositories
Access GatingOpen AccessUnrestricted download
Evaluation LicenseOpen AccessDataset usage and redistribution terms
Frontier API Compute Cost:
$5 – $20 USD for full benchmark evaluation run on frontier APIs.
Recommended Local GPU Setup:
1x NVIDIA RTX 4090 (24GB) or A100 (40GB/80GB) via vLLM / SGLang
Official Dataset & Benchmark Files:Download / View Dataset Repository ↗
How to Run & Reproduce
Standardized evaluation protocols, CLI commands, and reproducible runner templates.Prompt Regimezero-shot
Reasoning Modedirect
Sampling Temp0
Pass@k Budgetk = 1
Tools & SandboxPure Text
Scoring Verifierexact-match
Option AEleutherAI LM-Evaluation-Harness (Open-Weight Models)
lm_eval --model hf --model_args pretrained=<model_path> --tasks arc-agi-2 --batch_size autoOption BOpenCompass Evaluation Framework
opencompass --datasets arc-agi-2 --models <model_config>Python APIDeterministic Inference Loop Snippet
# 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,
)Standardized Reporting Requirement:
When publishing results for ARC-AGI-2, 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.Overall Contamination Risk:MEDIUM
Refresh Cadence:
Static fixed snapshot
Test Set Exposure:
Public on web / HuggingFace