Small Transformer Achieves 44% on ARC-AGI-1 in 1.5 Hours for 67 Cents.

An independent AI researcher has demonstrated that a small transformer trained from scratch can achieve 44 percent on the ARC-AGI benchmark while incurring merely 67 cents in compute costs. The model, completed in 1.5 hours on a single RTX 5090 graphics card, matches the performance of larger language models utilizing test-time training methodologies. The project, now open-sourced, underscores a significant shift toward highly efficient, sample-driven artificial intelligence development. The ARC-AGI benchmark, originally designed to evaluate fluid intelligence and abstract reasoning, has served as a rigorous testing ground for AI systems for six years. While large language models have historically struggled with its symbolic reasoning tasks, this new approach bypasses massive pretraining in favor of direct, supervised learning focused exclusively on output tokens. The architecture eliminates input token training during the loss calculation, relies on a NorMuon optimizer paired with an auxiliary AdamW scheduler, and incorporates specialized positional encodings such as three-dimensional rotary position embeddings and per-task dihedral embeddings. These architectural refinements, combined with carefully filtered training data drawn from ARC-1 and non-overlapping segments of ARC-2, yielded a 7 percent score on the newer ARC-2 subset without introducing data leakage. The developer emphasized that sample efficiency remains the most critical bottleneck in contemporary AI research. By restricting computation to modern deep learning frameworks and consumer-grade hardware, the project demonstrates that substantial reasoning capabilities can emerge without recursive algorithms, synthetic data generation, or excessive parameter scaling. Open-source code has been released to encourage further iteration, with the researcher projecting that performance could approach 65 percent through modest architectural adjustments. Independent validation from prominent AI scientists has highlighted the approach potential to lower experimental barriers and redirect focus toward fundamental representation learning rather than brute-force scaling. Despite the efficiency gains, the release has reignited discussions surrounding benchmark methodology and broader AI paradigms. Critiques regarding parameter inflation in competing test-time training approaches and the reliance of language models on post-training optimization have been addressed, reinforcing the position that transformer architectures retain untapped potential for general reasoning tasks. As the artificial intelligence community continues to evaluate fluid intelligence benchmarks, this low-cost implementation offers a viable alternative to resource-intensive development pipelines, emphasizing reproducibility, algorithmic transparency, and scalable experimentation.