Pathway BDH-CQ hits 29.2% pass@2, ARC-AGI cost record

The pursuit of AI has increasingly relied on scaling: more data, more parameters, and more computation. Pathway’s BD architecture takes a different route. Instead of generating extra tokens as a chain-of-thought, it refines a response in a hidden, latent space, which avoids the memory and compute costs of processing each reasoning step verbally.

This direction builds on AWS infrastructure, with Amazon SageMaker HyperPod for distributed training and EC2 p5en.48xlarge instances for compute. The practical payoff is a dramatic reduction in cost per task: at US $0.0007, BDH-CQ undercuts previous methods, making the technology viable for real-time, high-volume uses like cybersecurity analysis or logistics coordination where every query’s cost matters.

The test ahead is whether BDH’s principles hold up outside the ARC-AGI-1 grid puzzles. While the benchmark result is promising, success in production hinges on how well the architecture handles more diverse, unstructured problems without degrading performance or requiring specialized infrastructure. If it does, the industry’s focus on model size may shift toward models that are simply designed to think more efficiently.