Pathway 150M Model Breaks ARC-AGI-1 Cost Frontier with Latent Recurrent Reasoning

Pathway released a 150M-parameter model that achieves state-of-the-art cost-efficiency on ARC-AGI-1 by using latent recurrent reasoning instead of chain-of-thought. Validated by Łukasz Kaiser, this challenges the assumption that reasoning requires expensive token generation and highlights architecture innovation over scale. The model significantly undercuts GPT-5.6 Luna in cost.