Benchmark · reasoning

ARC-AGI 2

33 results 28 models

ARC-AGI 2 (Abstraction & Reasoning Corpus, version 2, run as the ARC Prize 2025) measures fluid reasoning: inferring an abstract rule from a few examples and applying it to a new case. The score is the percentage of puzzles solved correctly on a held-out, semi-private evaluation set.

Read more
Example
A task gives a few colored-grid input→output pairs (a small grid of colored cells changed in one consistent way), and the solver must work out the hidden transformation and produce the correct output grid for a new input.
Scoring
Each puzzle is pass/fail: the produced output grid is compared to the exact correct grid, and the reported score is the percentage of puzzles solved on the semi-private evaluation set.
Verification
Results are checked automatically by exact match on the output grid — the predicted grid must match the answer cell for cell — on a semi-private test set that is not publicly released to prevent memorization.
Why it matters
It focuses on puzzles that are easy for people but hard for AI, so it is a closely watched measure of real abstraction and general reasoning rather than memorized knowledge.
Worked example
Task
ARC-AGI-2 grid puzzle: infer the transformation from the training pairs, then give the output grid for the test input (digits 0–9 are colors, 0 = black background). Train 1: [[4,0,0,0],[0,0,6,0],[4,0,0,0],[0,0,6,0]] → [[0,0,0,0],[0,0,0,0],[4,0,6,0],[4,0,6,0]]. Train 2: [[0,3,0,0],[0,0,0,8],[0,3,0,0],[0,0,0,0]] → [[0,0,0,0],[0,0,0,0],[0,3,0,0],[0,3,0,8]]. Test: [[0,0,2,0],[5,0,0,0],[0,0,2,0],[5,0,0,0]] → ?
Solution
[[0, 0, 0, 0], [0, 0, 0, 0], [5, 0, 2, 0], [5, 0, 2, 0]]
Walkthrough
Every training pair shows one rule — 'gravity': each colored cell falls straight down to the lowest free cells of its column, keeping its color and count, while the 0-background moves to the top. On the test, column 0's two 5s and column 2's two 2s settle into the bottom two rows. ARC-AGI-2 grades by exact grid match — right dimensions and every cell value — scored pass@2 over two allowed attempts.
0 25 50 75 100 2024-12-20 2025-10-11 2026-08-03 o3 tuned high · 87 · 2024-12-20 public AI reasoning systems · 9 · 2025-03-24 o4-mini-medium · 3 · 2025-04-22 o4-mini-low · 3 · 2025-04-22 o3-medium · 3 · 2025-04-22 o3-low · 3 · 2025-04-22 Grok 4 · 15.9 · 2025-07-09 Grok 4 · 15.9 · 2025-08-08 HRM · 2 · 2025-08-15 Grok-4 · 29.4 · 2025-09-16 TRM · 8 · 2025-10-06 Gemini 3 Deep Think · 45.1 · 2025-11-18 Gemini 3 Deep Think · 45.1 · 2025-11-18 Gemini 3 Deep Think · 84.6 · 2026-02-12 Gemini 3 Pro (Poetiq refinement) · 54 · 2025-12-05 NVARC (fine-tuned 4B model variant) · 27.6 · 2025-12-05 Poetiq's system (Gemini-based configuration) · 54 · 2025-12-05 Opus 4.5 (Thinking, 64k) · 37.6 · 2025-12-05 Tiny Recursive Model (TRM) · 8 · 2025-12-05 GPT-5.2 Thinking · 52.9 · 2025-12-11 GPT-5.2 Thinking · 52.9 · 2025-12-11 Claude Opus 4.6 · 68.8 · 2026-02-05 Claude Opus 4.6 · 69 · 2026-04-16 Gemini 3.1 Pro · 77.1 · 2026-02-19 Imbue’s Darwinian Evolver · 95.1 · 2026-04-16 Confluence Lab · 97.9 · 2026-04-16 Gemini 3 “Deep Think” · 84.6 · 2026-04-16 GPT-5.4 Pro · 83.3 · 2026-04-16 GPT-5.2 · 53 · 2026-04-16 GPT-5.5 · 85 · 2026-04-25 Gemini 3 Pro · 54 · 2026-07-06 GPT-5.2 Pro · 54.2 · 2026-07-06 Inkling-Small · 40.1 · 2026-08-03
o3 tuned high public AI reasoning systems o4-mini-medium o4-mini-low o3-medium o3-low Grok 4 HRM Grok-4 TRM Gemini 3 Deep Think Gemini 3 Pro (Poetiq refinement) NVARC (fine-tuned 4B model variant) Poetiq's system (Gemini-based configuration) Opus 4.5 (Thinking, 64k) Tiny Recursive Model (TRM) GPT-5.2 Thinking Claude Opus 4.6 Gemini 3.1 Pro Imbue’s Darwinian Evolver Confluence Lab Gemini 3 “Deep Think” GPT-5.4 Pro GPT-5.2 GPT-5.5 Gemini 3 Pro GPT-5.2 Pro Inkling-Small
Timeline
Date Model Score Source
2026-08-03 Inkling-Small 40.1% Thinking Machines Lab releases Inkling-Small, a 276B open weights multimodal MoE model
2026-07-06 Gemini 3 Pro 54.0% Modality-Driven Search with Holistic Trace Judging for ARC-AGI-2
2026-07-06 GPT-5.2 Pro 54.2% Modality-Driven Search with Holistic Trace Judging for ARC-AGI-2
2026-04-25 GPT-5.5 85.0% OpenAI releases GPT-5.5, a ground-up retrained model with native omnimodal processing
2026-04-16 Imbue’s Darwinian Evolver 95.1% GPT-5.2 achieves ~53% on ARC-AGI-2 benchmark in December 2025
2026-04-16 Confluence Lab 97.9% GPT-5.2 achieves ~53% on ARC-AGI-2 benchmark in December 2025
2026-04-16 Gemini 3 “Deep Think” 84.6% GPT-5.2 achieves ~53% on ARC-AGI-2 benchmark in December 2025
2026-04-16 GPT-5.4 Pro 83.3% GPT-5.2 achieves ~53% on ARC-AGI-2 benchmark in December 2025
2026-04-16 GPT-5.2 53.0% GPT-5.2 achieves ~53% on ARC-AGI-2 benchmark in December 2025
2026-04-16 Claude Opus 4.6 69.0% GPT-5.2 achieves ~53% on ARC-AGI-2 benchmark in December 2025
2026-02-19 Gemini 3.1 Pro 77.1% Google releases Gemini 3.1 Pro with improved reasoning
2026-02-12 Gemini 3 Deep Think 84.6% Google releases upgraded Gemini 3 Deep Think with new benchmark scores
2026-02-05 Claude Opus 4.6 68.8% Anthropic releases Claude Opus 4.6 with 1M context window and agentic improvements
2025-12-11 GPT-5.2 Thinking 52.9% OpenAI introduces GPT-5.2 with improved coding, long-context, and reasoning capabilities
2025-12-11 GPT-5.2 Thinking 52.9% OpenAI releases GPT-5.2, outperforming Gemini 3 in coding and reasoning benchmarks
2025-12-05 Gemini 3 Pro (Poetiq refinement) 54.0% ARC Prize 2025 results highlight refinement loops as key to AGI progress
2025-12-05 NVARC (fine-tuned 4B model variant) 27.64% NVIDIA researchers win Kaggle ARC Prize 2025 with cost-efficient AGI reasoning
2025-12-05 Poetiq's system (Gemini-based configuration) 54.0% Poetiq achieves 54% on ARC-AGI-2 at half the cost using meta-system
2025-12-05 Opus 4.5 (Thinking, 64k) 37.6% ARC Prize 2025 results highlight refinement loops as key to AGI progress
2025-12-05 Tiny Recursive Model (TRM) 8.0% ARC Prize 2025 results highlight refinement loops as key to AGI progress
2025-11-18 Gemini 3 Deep Think 45.1% Google releases Gemini 3 Pro with state-of-the-art reasoning and multimodal capabilities
2025-11-18 Gemini 3 Deep Think 45.1% Google releases Gemini 3 Pro with state-of-the-art reasoning and multimodal capabilities
2025-10-06 TRM 8.0% Tiny Recursive Model (TRM) achieves higher ARC-AGI accuracy than LLMs with 7M parameters
2025-09-16 Grok-4 29.4% ARC-AGI v2 SoTA set by evolving English instructions with Grok-4
2025-08-15 HRM 2.0% Analysis finds HRM's ARC-AGI performance driven by outer loop refinement and memorization
2025-08-08 Grok 4 15.9% xAI Grok 4 leads Arc-AGI2 benchmark with 15.9% score versus GPT5
2025-07-09 Grok 4 15.9% xAI releases Grok 4 with scaled reinforcement learning and native tool use
2025-04-22 o4-mini-medium 3.0% ARC Prize Foundation evaluates OpenAI o3 and o4-mini on ARC-AGI benchmarks
2025-04-22 o4-mini-low 3.0% ARC Prize Foundation evaluates OpenAI o3 and o4-mini on ARC-AGI benchmarks
2025-04-22 o3-medium 3.0% ARC Prize Foundation evaluates OpenAI o3 and o4-mini on ARC-AGI benchmarks
2025-04-22 o3-low 3.0% ARC Prize Foundation evaluates OpenAI o3 and o4-mini on ARC-AGI benchmarks
2025-03-24 public AI reasoning systems 9.0% ARC Prize Foundation launches ARC-AGI-2 benchmark and ARC Prize 2025
2024-12-20 o3 tuned high 87.0% OpenAI previews o3 reasoning model with ARC AGI and Frontier Math breakthroughs