Skip to content
ModelCap

Benchmark leaderboard

ARC-AGI-3 leaderboard

ARC-AGI-3 moves the ARC challenge into interactive game-like environments where the model must learn the rules while acting. The verified board reports the percentage of tasks completed. This page lists every model ModelCap tracks with a published ARC-AGI-3 result, its best evaluated configuration, the source's own rank, and the model's ModelCap Index position, API price and context window.

Snapshot as of 18 November 2026

Claude Fable 5.2 from Anthropic leads the ARC-AGI-3 leaderboard among the 28 tracked models with 81.2% (X-High configuration), per the source snapshot published 18 November 2026.

Tracked models with a result
28
Entries on the source board
151
Open-weight models listed
3

ARC-AGI-3 standings (28 models)

Sorted by published score, one row per model. Source: ARC Prize Foundation.

ARC-AGI-3 leaderboard: every tracked model's published score, source rank, ModelCap Index position, price and context
#ModelScoreConfigurationSource rankModelCap IndexOutput $/1MContextWeightsPublished
1Claude Fable 5.2Anthropic81.2%X-High+4 more evaluated#32 / 151#6$40.001MNo linked weights18 November 2026
2GPT-6.2 AstraOpenAI70.3%Max+4 more evaluated#36 / 151#22$40.001MNo linked weights18 November 2026
3Gemini 4 CarbonGoogle65.1%Low+1 more evaluated#38 / 151#1$10.001MNo linked weights18 November 2026
4GPT-6 AstraOpenAI62.7%Max+5 more evaluated#40 / 151—$50.001MNo linked weights18 November 2026
5GPT-6.2 SolOpenAI54.9%Max+4 more evaluated#44 / 151#29$10.001MNo linked weights18 November 2026
6GPT-6.1 SolOpenAI52.7%Max+4 more evaluated#47 / 151—$10.001MNo linked weights18 November 2026
7Gemini 4 ArgonGoogle49.8%Low+1 more evaluated#49 / 151#3$10.001MNo linked weights18 November 2026
8Claude Opus 5Anthropic30.2%High#60 / 151—$25.001MNo linked weights18 November 2026
9Claude Sonnet 5.5Anthropic24.3%X-High+3 more evaluated#62 / 151#24$10.001MNo linked weights18 November 2026
10Gemini 3.8 FlashGoogle10.4%High+2 more evaluated#69 / 151#8$4.131MNo linked weights18 November 2026
11GPT-5.6 SolOpenAI7.8%Max+4 more evaluated#74 / 151—$20.001MNo linked weights18 November 2026
12GPT-6 SolOpenAI4.6%Max+5 more evaluated#82 / 151—$10.001MNo linked weights18 November 2026
13Grok 4.6xAI2.1%X-High#91 / 151—$6.00500KNo linked weights18 November 2026
14Grok 4.7xAI1.8%X-High+3 more evaluated#92 / 151#59$6.00500KNo linked weights18 November 2026
15Claude Opus 4.8Anthropic1.5%High#97 / 151—$25.001MNo linked weights18 November 2026
16DeepSeek V4.1 FlashDeepSeek0.9%Max+3 more evaluated#102 / 151—$1.201MOpen weights18 November 2026
17GPT-5.6 TerraOpenAI0.8%Max+4 more evaluated#103 / 151—$12.001MNo linked weights18 November 2026
18Claude Opus 4.6Anthropic0.5%Max#108 / 151—$25.001MNo linked weights18 November 2026
19Claude Haiku 5.5Anthropic0.5%Max#113 / 151#93$0.501MNo linked weights18 November 2026
20GPT-5.5OpenAI0.4%High#114 / 151#18$30.001MNo linked weights18 November 2026
21Grok 4.5xAI0.3%Medium+2 more evaluated#124 / 151—$6.00500KNo linked weights18 November 2026
22Inkling 2 SmallThinking Machines0.3%X-High#127 / 151#98$1.001MOpen weights18 November 2026
23DeepSeek V4.1 Flash 1017DeepSeek0.2%Max#132 / 151#92$0.601MOpen weights18 November 2026
24GPT-5.4OpenAI0.2%High#133 / 151—$15.001MNo linked weights18 November 2026
25GPT-6 LunaOpenAI0.2%Medium+5 more evaluated#134 / 151#58$0.501MNo linked weights18 November 2026
26Claude Opus 4.7Anthropic0.2%High#136 / 151—$25.001MNo linked weights18 November 2026
27GPT-5.6 LunaOpenAI0.2%Max+4 more evaluated#136 / 151—$1.201MNo linked weights18 November 2026
28Grok 4.20xAI0.1%Reasoning#145 / 151—$2.502MNo linked weights18 November 2026

How ModelCap uses ARC-AGI-3

ARC-AGI-3 ranks models by verified ARC-AGI-3 task completion. ModelCap ingests the board as published, matches each entry to a catalogue model with a reviewed identity, and shows the source's score, interval and rank unchanged. Not on its own. The ModelCap Index combines several public capability sources with published uncertainty; ARC-AGI-3 contributes as reasoning evidence where a model has an admitted result. The methodology page documents the weighting. Read the methodology and identity rules.

ARC-AGI-3 leaderboard: common questions

What is the ARC-AGI-3 benchmark?

ARC-AGI-3 moves the ARC challenge into interactive game-like environments where the model must learn the rules while acting. The verified board reports the percentage of tasks completed. It is published by ARC Prize Foundation.

Which AI model leads ARC-AGI-3 right now?

Claude Fable 5.2 (Anthropic) holds the top ARC-AGI-3 score among the models ModelCap tracks, at 81.2% as of the source snapshot published 18 November 2026.

How many models are ranked on the ARC-AGI-3 leaderboard here?

28 tracked models have a published ARC-AGI-3 result on ModelCap; the source board itself lists 151 entries. Each row shows the model's best evaluated configuration.

What is the best open-weight model on ARC-AGI-3?

DeepSeek V4.1 Flash from DeepSeek is the highest-scoring model with openly downloadable weights on this board, at 0.9%.

Which model offers the best value on ARC-AGI-3?

Among the ten highest-scoring priced models, Gemini 3.8 Flash has the lowest listed output price at $4.13 per 1M tokens while scoring 10.4%.

How is ARC-AGI-3 scored?

The board ranks models by verified ARC-AGI-3 task completion; higher scores are better. ModelCap shows the source's own score, interval and rank and never re-runs the evaluation.

Does ARC-AGI-3 decide the ModelCap Index rank?

Not on its own. The ModelCap Index combines several public capability sources with published uncertainty; ARC-AGI-3 contributes as reasoning evidence where a model has an admitted result. The methodology page documents the weighting.

How recent are the ARC-AGI-3 results?

The newest ARC-AGI-3 publication ModelCap holds is dated 18 November 2026. This page uses the published snapshot; source collection, publication and page updates happen separately.

Explore further