Pathway unveils efficient reasoning model challenging frontier model economics

Pathway announced BDH-CQ, a 150-million-parameter reasoning model using a post-transformer architecture, achieving 29.5% accuracy on the ARC-AGI-1 benchmark at a computed inference cost of $0.0007 per task. The model demonstrates comparable performance to OpenAI's GPT-5.6 Luna at a fraction of the cost, operating without chain-of-thought reasoning and requiring less training data. The breakthrough challenges the current economics of frontier models, suggesting that architectural innovation could deliver significantly more efficient alternatives to large-scale model approaches.