Pathway's 150M-parameter BDH-CQ hits 29.2% pass@2 on ARC-AGI-1 at $0.0007 per task

AWS Machine Learning Blog (Pathway)·high signalXBlueskyLinkedInCopy link

Baby Dragon Hatchling is a post-transformer architecture that iterates in a recurrent latent state instead of emitting chain-of-thought tokens, with about 5% of neurons active at a time and internal memory updated during inference rather than by fine-tuning. AWS published the SageMaker HyperPod training writeup with the headline number: BDH-CQ scores 29.2% pass@2 on ARC-AGI-1 at $0.0007 per task with 150 million parameters. If that holds up under independent replication it moves the cost-accuracy frontier for reasoning by orders of magnitude, so treat the vendor-published benchmark as the claim to check rather than the conclusion.