Tiny reasoning model breaks cost-accuracy Pareto frontier on Arc-AGI

dank_philosopher · reddit · 2026-08-17

AI lab Pathway released benchmark results for BDH-CQ, a 150M parameter reasoning model. It scored 29.5% pass@2 on the ARC-AGI 1 public set at an inference cost of $0.0007 per task, making it roughly 11x cheaper per task than GPT 5.6 Luna. While Luna scored slightly higher (34.2%), BDH-CQ demonstrates a significant breakthrough in cost-efficiency. The approach combines in-context learning with recurrent latent reasoning without verbalizing intermediate results, with pretraining experiments scaling from 1B to 600B parameters.

Related event: Pathway's Tiny Model Sets ARC-AGI Cost-Efficiency Record(2 posts)→

Original post →

More from Models

Models channel →