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)→
More from Models
- 前沿智能降价如云存储,企业 AI 竞争转向路由与编排层 — krishnan · 2026-08-18
- DeepSeek-V4-Pro Released with Adjustable Reasoning and OpenAI API Compatibility — petrusenko_max · 2026-08-18
- Microsoft Research: 4B model tuned with SocialRL out-negotiates GPT-5 family — dair_ai · 2026-08-17
- Gemini 3.7 Flash Launches with Temporary Low Pricing — KoseteBamse · 2026-08-17
- New Flash model matches Pro in style and performance — teortaxesTex · 2026-08-17
- GLM-5.3 outperforms Fable 5 in 3D web design with 15x lower cost — togethercompute · 2026-08-17