Pathway's post-transformer AI promises 11x cheaper reasoning than leaders

Pathway researchers debut BDH-CQ, a post-transformer AI with a latent reasoning engine and vector-based memory that dramatically reduces the memory and compute costs of reasoning versus traditional transformer models. In ARC-AGI-1 benchmarks, BDH-CQ solved 3 of 10 puzzles in two attempts, using only 150 million parameters—far fewer than frontier models like Llama 3—while achieving substantially lower per-token costs (about 11x cheaper than OpenAI-style baselines). Pathway plans to scale BDH-CQ to up to 600 billion parameters and apply the approach to complex tasks in cybersecurity and industrial operations, with independent researchers verifying the ARC-AGI-1 results and signaling a potential new direction for efficient AI reasoning.