The ARC-AGI leaderboard shows model performance on three versions of the benchmark, measuring fluid intelligence and efficient adaptation, with trend lines for reasoning systems and raw LLMs.
# ARC Prize - Leaderboard
Source: [https://arcprize.org/leaderboard](https://arcprize.org/leaderboard)
ARC\-AGI\-1
ARC\-AGI\-2
ARC\-AGI\-3
### Understanding the Leaderboard
ARC\-AGI has evolved from its first versions \(ARC\-AGI\-1 and 2\) which measured passive fluid intelligence, to ARC\-AGI\-3 which challenges AI agents to adapt on the fly to novel interactive environments\.
The scatter plot above visualizes the critical relationship between cost\-per\-task and performance \- a key measure of efficiency\. True intelligence isn't just about solving problems, but solving them efficiently with minimal resources\.
#### Interpreting the data
- **Reasoning Systems Trend Line**solutions display connected points representing the same model at different reasoning levels\. These trend lines illustrate how increased reasoning time affects performance, typically showing asymptotic behavior as thinking time increases\.
- **Base LLMs**solutions represent single\-shot inference from standard language models like GPT\-4\.5 and Claude 3\.7, without extended reasoning capabilities\. These points demonstrate raw model performance without additional reasoning enhancements\.
- **Kaggle Systems**solutions showcase competition\-grade submissions from the Kaggle challenge, operating under strict computational constraints \($50 compute budget for 120 evaluation tasks\)\. These represent purpose\-built, efficient methods specifically designed for the ARC Prize\.
#### Verification Policy
For more information, see our[testing policy](https://arcprize.org/policy)\.
### Leaderboard Breakdown
#### Notes
Only systems which required less than $10,000 to run are shown\.
For models that were not able to produce full test out puts, remaining tasks were marked as incorrect\.
Results marked as "preview" are unofficial and may be based on incomplete testing\.
1ARC\-AGI\-2 score estimate based on partial testing results and o1\-pro pricing\.
2Provisional cost estimates based on Gemini 3 Pro pricing\. Model to be retested once released\.
GPT-6 Sol and Astra have taken the top spots on the ARC-AGI-3 leaderboard, marking a notable advance in abstract reasoning benchmarks for frontier AI models.
Top Kaggle scores on the ARC-AGI-3 benchmark jumped from 7% to 56% in just 30 days, with small local models in a harness beating average humans on a benchmark designed to demonstrate human cognitive superiority.
GPT-6 Astra by OpenAI achieves state-of-the-art performance on the ARC-AGI-3 benchmark, scoring 63% and surpassing human performance on 96% of levels, demonstrating advanced symbolic modeling capabilities.
A new benchmark called LisanBench evaluates LLMs on word chain tasks requiring planning, memory, and constraint adherence, with results showing strong performance from o3 and Anthropic models.
A critique arguing that ARC-AGI 3 unfairly disables an agent's ability to maintain context across actions, making it an dishonest measure of general intelligence. It notes that allowing compaction triples scores while using far fewer tokens, and that real-world agents work that way.