AI Reasoning Leaps: Cheaper, Smarter Models Emerge

Pathway, an AI lab specializing in Post-Transformer architectures, has announced impressive benchmark results for its BDH-CQ reasoning model. The 150M-parameter model attained a 29.5% pass@2 score on the ARC-AGI-1 evaluation set.

This accomplishment was achieved at a remarkably low inference cost of $0.0007 per task. This price is approximately eleven times less expensive than the comparable reasoning models used by ChatGPT.

The cost savings stem from BDH-CQ's unique architecture. Unlike many AI systems that generate extensive intermediate text, BDH-CQ performs reasoning internally within its memory, significantly reducing computational overhead.

Amazon Web Services views this development as a critical step toward deploying advanced AI reasoning in production environments cost-effectively. This efficiency could redefine the landscape of AI capabilities versus cost.

While OpenAI's Luna model scored slightly higher at 34.2%, its cost per task is substantially more at $0.008, even after recent price reductions. Other frontier models like Claude Opus 5 and Gemini 3.1 Pro achieve higher scores but cost hundreds of times more per task.

Pathway's findings suggest that model architecture, in addition to scale, is a key driver for the next leap in AI reasoning. The company plans to apply this efficient approach to more challenging benchmarks, including mathematical reasoning and ARC-AGI-2 and -3 evaluations.