Pathway tests a lower-cost model architecture

Pathway unveiled BDH-CQ, a 150-million parameter small reasoning model built on a post-transformer architecture that the company says can deliver comparable performance with far lower computing resources than leading frontier models. The model is designed to reason without chain-of-thought and rely on improved memory to reduce data needs.

On the ARC-AGI-1 benchmark, BDH-CQ achieved 29.5% pass@2 accuracy with a computed inference cost of $0.0007 per task. OpenAI’s GPT 5.6 Luna (Low), whose price was reduced by 80% on July 30, scored 34.5% on the same benchmark, but still cost 11 times more than Pathway’s model.

CEO and co-founder Zuzanna Stamirowska described the work as a bid to “squeeze more intelligence per dollar” by changing the underlying paradigm. Pathway has framed the breakthrough as “a PageRank moment for intelligence,” and its team includes advisors and executives with backgrounds at Google DeepMind, Databricks, NYU and early Google.

Pathway first outlined its approach in October 2025 with “The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain.” The company believes the architecture can scale to 600B parameter models, while the wider neolab sector includes over 40 companies that have raised $40B to pursue new AI efficiency breakthroughs.