The ARC Prize 2025 competition concluded with a top Kaggle score of 24% on the ARC-AGI-2 private dataset, while commercial models and bespoke refinement solutions achieved significantly higher performance. The event featured 1,455 teams submitting over 15,000 entries and 90 research papers, all of which are open-source.
- Top Kaggle winner reached 24% accuracy for $0.20/task; top verified commercial model Opus 4.5 scored 37.6% for $2.20/task.
- Poetiq’s refinement solution built on Gemini 3 Pro achieved 54% accuracy for $30/task, up from a baseline of 31%.
- Tiny Recursive Model (TRM) achieved 45% test-accuracy on ARC-AGI-1 with only 7M parameters using recursive improvement.
- CompressARC introduced 76K parameters and achieved 20% on ARC-AGI-1 by minimizing description length at test time without pretraining.
The analysis identifies iterative refinement loops as the central theme driving AGI progress, enabling models to incrementally optimize programs based on feedback signals rather than relying solely on static neural networks.