Inference
EN
ARC Prize Foundation

ARC-AGI-3 (semi-private)

Abstract reasoning puzzles held out from training data.

ModelsModels with a result.26
Top resultBest result on this benchmark.65.7%Claude Fable 5.2
Top-3 spreadPoints from first to third.3.0 pts
YearYear of release.—

Result and price

25 of 26 results

Results

26 results
#ModelResultThe score from the source.
01Claude Fable 5.2max effort65.7%
02Gemini 4 Carbonmedium effort63.9%
03GPT-6 Astramax effort62.7%
04GPT-6.1 Solmax effort52.7%
05Gemini 4 Argonmedium effort46.3%
06Claude Opus 5high effort30.2%
07Claude Sonnet 5.5xhigh effort24.3%
08Gemini 3.8 Flashhigh effort10.4%
09GPT-5.6 Solmax effort7.8%
10GPT-6 Solmax effort4.6%
11Grok 4.6xhigh effort2.1%
12Grok 4.7xhigh effort1.8%
13Claude Opus 4.8high effort1.5%
14DeepSeek V4.1 Flash 1017max effort0.9%
15GPT-5.6 Terramax effort0.8%
16Claude Opus 4.6max effort0.5%
16Claude Haiku 5.5max effort0.5%
18GPT-5.5high effort0.4%
18Gemini 3.1 Pro Preview0.4%
20Grok 4.5high effort0.3%
21GPT-5.4high effort0.2%
21GPT-5.6 Lunamax effort0.2%
21Claude Opus 4.7high effort0.2%
24Muse Spark Openxhigh effort0.1%
24GPT-6 Lunamax effort0.1%
24Grok 4.20 (Beta Reasoning)0.1%