Researchers at AI company Pathway have unveiled a novel artificial intelligence model called BDH-CQ that significantly reduces operational expenses compared to existing systems. According to a research paper published on preprint server arXiv, the new model operates at roughly one-eleventh the cost of OpenAI’s entry-level reasoning model while achieving competitive performance on cognitive benchmarks.
The BDH-CQ model employs a fundamentally different architecture from mainstream AI systems. Rather than relying on traditional transformer-based designs used by ChatGPT and Claude, Pathway’s approach utilizes numerical arrays and vectors to represent abstract reasoning patterns. This innovation allows the system to solve complex puzzles and reasoning tasks without accumulating the memory demands that plague conventional models as they process increasingly lengthy inputs.
When evaluated on the ARC-AGI benchmark—a standardized test measuring artificial intelligence progress through nonverbal reasoning challenges—BDH-CQ achieved approximately 30 percent accuracy using just 150 million parameters. Though this score trails some competing models, the dramatic cost efficiency represents a significant breakthrough. Researchers suggest that expanding the architecture to larger parameter sizes could substantially enhance reasoning capabilities while maintaining affordability advantages.
Pathway intends to scale the technology further and develop applications targeting sectors including cybersecurity and industrial operations. Independent verification by prominent AI researchers, including a co-author of the foundational transformer paper, has validated the model’s results and performance claims.
Originally reported by Livescience →
About the Author
