150M-parameter model scores 29.2% on ARC-AGI-1

Pathway's Baby Dragon Hatchling, a 150-million-parameter model, scored 29.2% pass@2 on ARC-AGI-1 at roughly $0.0007 per task, according to AWS's SageMaker HyperPod writeup on the project. That's a fraction of the parameter count and per-task cost most reasoning models run at.

The architecture, called BDH-CQ, skips the usual move of writing out chain-of-thought tokens. Instead it iterates in a recurrent latent state, with about 5% of neurons active at any point, and updates its internal memory during inference rather than through a separate fine-tuning pass. If that holds up, it's a different way of getting a model to reason step by step: state that evolves as it runs, not a transcript of tokens it has to generate and re-read.

Why it matters: ARC-AGI-1 was built specifically to resist the kind of pattern memorization that lets big models look smart on benchmarks without generalizing. A small model doing well on it, if the number holds, says something about efficient reasoning that parameter count alone doesn't capture.

The catch is that this is AWS's own blog post about a partner's model trained on AWS infrastructure. It's a vendor-published benchmark, not a third-party eval, and the post doesn't say whether anyone outside Pathway and AWS has reproduced the score. I'd treat 29.2% as the number to go verify, not the number to build a roadmap around. If independent labs run BDH-CQ against ARC-AGI-1 and land anywhere close to that figure, the cost-per-task math changes by orders of magnitude for anyone running reasoning workloads at scale.