The honest view from inside technical leadership
LLM Leaderboard · Interactive reasoning

ARC-AGI-3 leaderboard

Interactive game-like environments the model has to explore, learn the rules of and solve, scored on the standard (no provider adapter) track.

← All models

Rankings

Snapshot 22 November 2026 · Scores copied from each benchmark's own leaderboard · How the CTO Index works

Score and price

#ModelProviderARC-AGI-3Price $/M
1Claude Fable 5.2Anthropic81.2%$8 / $40
2GPT-6.2 AstraOpenAI70.3%$8 / $40
3Gemini 4 CarbonGoogle65.1%$2 / $10
4GPT-6 AstraOpenAI62.7%$10 / $50
5GPT-6.2 SolOpenAI54.9%$2 / $10
6GPT-6.1 SolOpenAI52.7%$2 / $10
7Gemini 4 ArgonGoogle49.8%$2 / $10
8Claude Opus 5Anthropic30.2%$5 / $25
9Claude Sonnet 5.5Anthropic24.3%$2 / $10
10Gemini 3.8 FlashGoogle10.4%$0.75 / $3.75
11GPT-5.6 SolOpenAI7.8%$4 / $20
12GPT-6 SolOpenAI4.6%$2 / $10
13Grok 4.6xAI2.1%$1.60 / $4.80
14Grok 4.7xAI1.8%$2 / $6
15Claude Opus 4.8Anthropic1.5%$5 / $25
16Claude Haiku 5.5Anthropic0.5%$0.10 / $0.50
17Inkling 2 SmallThinking Machines0.3%$0.30 / $1
18DeepSeek V4.1 FlashDeepSeek0.2%$0.15 / $0.60
19GPT-6 LunaOpenAI0.2%$0.10 / $0.50

Other benchmarks

Frequently asked questions

What does ARC-AGI-3 measure?

Interactive game-like environments the model has to explore, learn the rules of and solve, scored on the standard (no provider adapter) track.

Which model has the highest ARC-AGI-3 score?

Claude Fable 5.2 from Anthropic leads ARC-AGI-3 at 81.2%, ahead of GPT-6.2 Astra (70.3%).

Why does ARC-AGI-3 matter to engineering leaders?

It tests learning on the job inside an environment, which is what long-running autonomous agents have to do.

What are the limitations of ARC-AGI-3?

Only a handful of frontier models have been run, and costs run into thousands of dollars per evaluation, so we show it but leave it out of the CTO Index.