ARC-AGI-2
Integrity rank #4 of 65 · 84 models scored · top score 95 · GPT-6 Astra
Fluid intelligence / novel skill via abstract grid puzzles
Strongest on discrimination, weakest on freshness. Caveat: the top scores are clustering near the ceiling.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | GPT-6 Astra | 95 | unverifiedT2 | 2026-09-03 |
| 2 | GPT-6.1 Sol | 94.17 | unverifiedT2 | 2026-09-29 |
| 3 | Claude Opus 5.5 | 93.33 | unverifiedT2 | 2026-09-22 |
| 4 | GPT-5.6 Sol | 92.5 | unverifiedT2 | 2026-07-09 |
| 5 | Claude Opus 5 | 90.42 | unverifiedT2 | 2026-07-24 |
| 6 | Claude Fable 5.1 | 90 | unverifiedT2 | 2026-09-01 |
| 7 | GPT-6 Sol | 89.58 | unverifiedT2 | 2026-09-22 |
| 8 | Gemini 3.8 Flash | 89.17 | unverifiedT2 | 2026-09-02 |
| 9 | Claude Fable 5 | 89.17 | unverifiedT2 | 2026-06-09 |
| 10 | GPT-5.5 | 85 | unverifiedT2 | 2026-04-23 |
| 11 | Gemini 3.7 Flash | 84.58 | unverifiedT2 | 2026-08-13 |
| 12 | GPT-5.5 Pro | 84.58 | unverifiedT2 | 2026-04-23 |
| 13 | GPT-5.6 Terra | 83.9 | unverifiedT2 | 2026-07-09 |
| 14 | GPT-5.4 Pro | 83.33 | unverifiedT2 | 2026-03-05 |
| 15 | Gemini 3.1 Pro | 77.1 | unverifiedT2 | 2026-02-19 |
| 16 | Claude Opus 4.7 | 75.83 | unverifiedT2 | 2026-04-16 |
| 17 | GPT-5.4 | 73.95 | unverifiedT2 | 2026-03-05 |
| 18 | DeepSeek V4.1 Flash | 72.92 | unverifiedT2 | 2026-09-09 |
| 19 | Gemini 3.5 Flash | 72.08 | unverifiedT2 | 2026-05-19 |
| 20 | Claude Opus 4.8 | 72.08 | unverifiedT2 | 2026-05-28 |
| 21 | Claude Opus 4.6 | 69.17 | unverifiedT2 | 2026-02-05 |
| 22 | Grok 4.6 | 67.08 | unverifiedT2 | 2026-08-12 |
| 23 | GLM-5.3-Flash | 65.83 | unverifiedT2 | 2026-08-20 |
| 24 | Grok 4.20 | 65.14 | unverifiedT2 | 2026-02-17 |
| 25 | Grok 4.7 | 61.39 | unverifiedT2 | 2026-09-21 |
| 26 | DeepSeek V4 Flash 0731 | 61.39 | unverifiedT2 | 2026-07-31 |
| 27 | DeepSeek V4 Pro 0813 | 61.25 | unverifiedT2 | 2026-08-13 |
| 28 | Gemini 3.6 Flash | 60.42 | unverifiedT2 | 2026-07-21 |
| 29 | Kimi K3 | 60.42 | unverifiedT2 | 2026-07-16 |
| 30 | Claude Sonnet 4.6 | 60.42 | unverifiedT2 | 2026-02-17 |
| 31 | GPT-5.6 Luna | 59.54 | unverifiedT2 | 2026-07-09 |
| 32 | GPT-6 Luna | 59.31 | unverifiedT2 | 2026-09-22 |
| 33 | GPT-5.2 Pro | 54.16 | unverifiedT2 | 2025-12-11 |
| 34 | GPT-5.2 | 52.91 | unverifiedT2 | 2025-12-11 |
| 35 | Grok 4.5 | 52.64 | unverifiedT2 | 2026-07-08 |
| 36 | Qwen 3.8 27B | 42.36 | unverifiedT2 | 2026-08-14 |
| 37 | Inkling-Small | 40.14 | unverifiedT2 | 2026-07-15 |
| 38 | Claude Opus 4.5 | 37.64 | unverifiedT2 | 2025-11-24 |
| 39 | Inkling | 36.53 | unverifiedT2 | 2026-07-15 |
| 40 | Gemini 3 Flash | 33.61 | unverifiedT2 | 2025-12-17 |
| 41 | Gemini 3 Pro | 31.11 | unverifiedT2 | 2025-11-18 |
| 42 | GLM-5.2 | 22.78 | unverifiedT2 | 2026-06-16 |
| 43 | GPT-5.4 Mini | 18.9 | unverifiedT2 | 2026-03-17 |
| 44 | GPT-5 Pro | 18.33 | unverifiedT2 | 2025-10-07 |
| 45 | GPT-5.1 | 17.64 | unverifiedT2 | 2025-11-13 |
| 46 | Grok 4 | 15.97 | unverifiedT2 | 2025-07-09 |
| 47 | Claude Sonnet 4.5 | 13.61 | unverifiedT2 | 2025-09-29 |
| 48 | Kimi K2.5 | 11.81 | unverifiedT2 | 2026-01-27 |
| 49 | Gemini 3.5 Flash-Lite | 10.28 | unverifiedT2 | 2026-07-21 |
| 50 | GPT-5 | 9.86 | unverifiedT2 | 2025-08-07 |
| 51 | Claude Opus 4 | 8.61 | unverifiedT2 | 2025-05-22 |
| 52 | o3 | 6.53 | unverifiedT2 | 2025-04-16 |
| 53 | o4-mini | 6.11 | unverifiedT2 | 2025-04-16 |
| 54 | Claude Sonnet 4 | 5.93 | unverifiedT2 | 2025-05-22 |
| 55 | GPT-5.4 Nano | 5.69 | unverifiedT2 | 2026-03-17 |
| 56 | Grok 4 Fast | 5.28 | unverifiedT2 | 2025-09-19 |
| 57 | MiniMax-M2.5 | 4.86 | unverifiedT2 | 2026-02-12 |
| 58 | GLM-5 | 4.86 | unverifiedT2 | 2026-02-11 |
| 59 | Gemini 2.5 Pro (Jun 2025) | 4.86 | unverifiedT2 | 2025-06-05 |
| 60 | o3-pro | 4.86 | unverifiedT2 | 2025-06-10 |
| 61 | GPT-5 mini | 4.44 | unverifiedT2 | 2025-08-07 |
| 62 | DeepSeek-V3.2 | 4.03 | unverifiedT2 | 2025-12-01 |
| 63 | Claude Haiku 4.5 | 4.03 | unverifiedT2 | 2025-10-15 |
| 64 | o3-mini | 2.99 | unverifiedT2 | 2025-01-31 |
| 65 | GPT-5 nano | 2.61 | unverifiedT2 | 2025-08-07 |
| 66 | Gemini 2.5 Flash (May 2025) | 2.54 | unverifiedT2 | 2025-05-20 |
| 67 | DeepSeek-R1 | 1.3 | unverifiedT2 | 2025-01-20 |
| 68 | Gemini 2.0 Flash (Feb 2025) | 1.3 | unverifiedT2 | 2025-02-05 |
| 69 | Qwen3-235B-A22B-Instruct (Jul 2025) | 1.25 | unverifiedT2 | 2025-07-25 |
| 70 | DeepSeek-R1 (May 2025) | 1.12 | unverifiedT2 | 2025-05-28 |
| 71 | Claude 3.7 Sonnet | 0.9 | unverifiedT2 | 2025-02-24 |
| 72 | o1-mini | 0.83 | unverifiedT2 | 2024-09-12 |
| 73 | Gemini 1.5 Pro (Sept 2024) | 0.8 | unverifiedT2 | 2024-09-24 |
| 74 | GPT-4.5 | 0.8 | unverifiedT2 | 2025-02-27 |
| 75 | Grok-3 mini | 0.42 | unverifiedT2 | 2025-06-24 |
| 76 | GPT-4.1 | 0.42 | unverifiedT2 | 2025-04-14 |
| 77 | Grok 3 | 0 | unverifiedT2 | 2025-04-09 |
| 78 | Llama 4 Maverick | 0 | unverifiedT2 | 2025-04-05 |
| 79 | Llama 4 Scout | 0 | unverifiedT2 | 2025-04-05 |
| 80 | Magistral Small 1.0 | 0 | unverifiedT2 | 2025-06-10 |
| 81 | GPT-4.1 mini | 0 | unverifiedT2 | 2025-04-14 |
| 82 | GPT-4.1 nano | 0 | unverifiedT2 | 2025-04-14 |
| 83 | GPT-4o (Nov 2024) | 0 | unverifiedT2 | 2024-11-20 |
| 84 | GPT-4o mini | 0 | unverifiedT2 | 2024-07-18 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark scores 84 models, and the spread between its best and worst recorded scores is 95.0 points.
2 cited facts
Its Benchmark Integrity Index score is 98, with grade A and rank 4 out of 65 tracked benchmarks. Age is the lowest-scoring component, at 88.7.
6 cited facts
It is flagged as saturated, with a top score of 95.0 and 5.0 points of headroom. The computed top-score cluster contains three models.
5 cited facts
The harness is classified as consistent, and the test set is held out. The evaluation set has not been released publicly; the record describes contamination risk as low by design. The harness notes describe extreme sensitivity to test-time compute and prize limits on compute and cost.
5 cited facts
This benchmark measures fluid intelligence and novel skill through abstract grid puzzles, with tasks designed by hand and calibrated against human performance, split between a public training and evaluation set and a secret evaluation set. Its design assumes the tasks are novel to any training corpus, and that puzzles humans solve readily while machine search finds hard serve to isolate general reasoning. The listed limitations are the narrow grid domain, program-search brute force, and the effect of search and compute budgets on scores.
4 cited facts
ARC-AGI-2: ARC-AGI-2 — the second-generation Abstraction and Reasoning Corpus of novel visual-grid puzzles measuring fluid, few-shot abstraction that resists brute-force memorization; scored as the percent of tasks solved. The evaluation set is held out (private) to keep it contamination-resistant.
GPT-6 Astra leads ARC-AGI-2 at 95. The full leaderboard above lists every recorded measurement, not just the headline number.
84 models have recorded scores on ARC-AGI-2, spanning a score spread of 95.
tensor.news grades ARC-AGI-2 A for integrity (score 98/100), ranking #4 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
Yes — top scores are clustering near the ceiling (saturation ratio 0.04), so ARC-AGI-2 no longer separates leading models well. Treat small gaps at the top with caution.
Its test set is held-out, which lowers — but does not eliminate — the risk that training data leaked into the questions.
Scores are reported under a consistent harness, so comparisons on ARC-AGI-2 are reasonably apples-to-apples.
Every ARC-AGI-2 measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites arxiv.org/abs/2505.11831.