A The ARC Atlas

Appendix E · Result registry

Every score, with its protocol attached.

A score alone does not identify an evaluation. Each row keeps it attached to the benchmark edition, system configuration, protocol or split, date, available cost or placement, and supporting evidence.

25
Evaluations
22
Machine results
3
Human baselines
4
With cost data

ARC-AGI-1

13 protocol-labelled evaluations.

Every row here belongs to ARC-AGI-1. Read the benchmark chapter →

SystemScoreProtocolDateCost or rankEvidence
Human baselineHuman baseline · arc1-human 98% Original ARC human baseline 2019 Not recorded On the Measure of Intelligence ↗
IcecuberResult · arc1-icecuber-2020 ~20% 2020 Kaggle solo private result 2020 Rank 1 Kaggle Abstraction and Reasoning Challenge ↗
GreenblattResult · arc1-greenblatt-2024 ~42% ARC-AGI-Pub 2024-06 Not recorded ARC Prize 2024 Winners & Technical Report ↗
GreenblattResult · arc1-greenblatt-public-2024 50% Public evaluation 2024-06 Not recorded ARC Prize 2024 Winners & Technical Report ↗
MIT Test-Time TrainingResult · arc1-mit-ttt-2024 53% Public validation, 8B model 2024-11 Not recorded The Surprising Effectiveness of Test-Time Training for ARC ↗
MIT Test-Time Training ensembleResult · arc1-mit-ttt-ensemble-2024 61.9% Ensemble with program synthesis 2024-11 Not recorded The Surprising Effectiveness of Test-Time Training for ARC ↗
The ARChitectsResult · arc1-architects-2024 53.5% 2024 Kaggle private evaluation 2024 Not recorded ARC Prize 2024 Winners & Technical Report ↗
The ARChitectsResult · arc1-architects-public-2024 71.6% Public evaluation 2024 Not recorded ARC Prize 2024 Winners & Technical Report ↗
MindsAIResult · arc1-mindsai-2024 55.5% 2024 season high 2024 Not recorded ARC Prize 2024 Winners & Technical Report ↗
BermanResult · arc1-berman-2024 53.6% ARC-AGI-Pub 2024-12 Not recorded How I got a record 53.6% on ARC-AGI-Pub ↗
OpenAI o3 (low)Result · arc1-o3-low-2024 75.7% 100-task semi-private evaluation, high efficiency 2024-12 Not recorded OpenAI o3 ARC-AGI breakthrough ↗
OpenAI o3 (high)Result · arc1-o3-high-2024 87.5% 100-task semi-private evaluation, approximately 172x compute 2024-12 Not recorded OpenAI o3 ARC-AGI breakthrough ↗
Tiny Recursive ModelResult · arc1-trm-2025 ~45% Reported ARC-AGI-1 result 2025 Not recorded ARC Prize 2025 Results & Analysis ↗

ARC-AGI-2

7 protocol-labelled evaluations.

Every row here belongs to ARC-AGI-2. Read the benchmark chapter →

SystemScoreProtocolDateCost or rankEvidence
Human baselineHuman baseline · arc2-human ~100% Every evaluation task solved by at least two humans 2025 Not recorded ARC-AGI-2 ↗
Frontier reasoning modelsResult · arc2-frontier-2025 ~1–3% Harness-free frontier-model range 2025 Not recorded ARC-AGI-2 ↗
NVARCResult · arc2-nvarc-2025 24.03% 2025 Kaggle private evaluation; contest constraints 2025-12-05 $0.2/task Rank 1 ARC Prize 2025 Results & Analysis ↗
Tiny Recursive ModelResult · arc2-trm-2025 ~8% Reported ARC-AGI-2 result 2025 Not recorded ARC Prize 2025 Results & Analysis ↗
Gemini 3 ProResult · arc2-gemini3-pro-2025 31% ARC Prize verified commercial baseline 2025-12-05 $0.81/task ARC Prize 2025 Results & Analysis ↗
Opus 4.5Result · arc2-opus45-2025 37.6% Thinking, 64k; ARC Prize verified 2025-12-05 $2.2/task ARC Prize 2025 Results & Analysis ↗
Poetiq + Gemini 3 ProResult · arc2-poetiq-gemini3-2025 54% Open refinement harness; ARC Prize verified 2025-12-05 $30/task ARC Prize 2025 Results & Analysis ↗

ARC-AGI-3

5 protocol-labelled evaluations.

Every row here belongs to ARC-AGI-3. Read the benchmark chapter →

SystemScoreProtocolDateCost or rankEvidence
Human baselineHuman baseline · arc3-human 100% First-contact human calibration 2026 Not recorded ARC-AGI-3 ↗
StochasticGooseResult · arc3-stochasticgoose-preview ~12.58% 30-day Preview Challenge 2025-08 Rank 1 ARC-AGI-3 Preview: 30-Day Learnings ↗
Blind SquirrelResult · arc3-blind-squirrel-preview ~6.71% 30-day Preview Challenge 2025-08 Rank 2 ARC-AGI-3 Preview: 30-Day Learnings ↗
Just-ExploreResult · arc3-just-explore-preview ~3.64% 30-day Preview Challenge private leaderboard 2025-08 Rank 3 ARC-AGI-3 Preview: 30-Day Learnings ↗
Frontier LLMsResult · arc3-frontier-harness-free ~0.1–0.5% Harness-free frontier 2025 Not recorded ARC-AGI-3 ↗

Evaluation receipts

The score is only one field.

A trustworthy result keeps the split, protocol, resource disclosure, artifacts, and claim boundary beside the number. These new ARC-AGI-3 records show the format explicitly.

ClaimResultSplitProtocolCostArtifactsBoundary
Human public corpus145 solves / 342 plays25 Public DemoFirst-run, one attempt90-minute sessionsStep replaysHuman cohort; not AI performance
EWM · GPT-5.5 high58.12% mean per-game RHAE25 public gamesFresh agent and workspace per gameNot normalized hereCode and run artifactsPrivate validation untested
OPINE-World78.4 action efficiency25 public gamesNo per-game trainingNot normalized herePaper-linked implementation evidencePublic set only
Verified EWM · gpt-5.6-sol~99% RHAE25 public gamesExploratory follow-upHigher-resource verificationAblation reportPublic-set saturation; model postdates games

YouTube first · local synths follow

ARC Radio

01 / 12 🦉 8-Bit Chiptune Playlist 🦉 Retro Video Game Music for Nostalgic Vibes YouTube · external stream

The 4 requested YouTube selections play first and require a network connection; their titles refresh from YouTube when they load. 8 original AI-composed retro-game loops follow and are generated live in your browser. Audio keeps playing when you close this panel and stops only when you press Pause.

Field notes · reader review

Help improve this guide

Found a wrong score, broken link, missing paper, or unclear passage? Tell us what you noticed.

How useful is it? optional
- / 5
What kind of note? optional

No account, no tracking. Sent straight to the maintainer.