AI Reasoning Cost Slashed: New Model Beats GPT‑5.6 Luna While Using a Fraction of Tokens

Pathway unveiled a post‑transformer AI model dubbed BDH‑CQ that redefines AI reasoning cost. In benchmark ARC‑AGI‑1 the system solved ↑30% of puzzles, yet its token consumption was ↓11× that of OpenAI’s GPT‑5.6 Luna.

AI reasoning cost breakthrough with BDH‑CQ

The 150‑million‑parameter architecture replaces textual logs with numerical vectors, enabling iterative “latent reasoning” loops that do not expand memory usage. Researchers say the design could scale to billions of parameters without the exponential token bill that plagues transformer models.

Why transformers strain GPUs

Traditional transformers revisit every prior token for each new word, a quadratic cost that inflates GPU memory as prompts grow. BDH‑CQ’s vector‑based workspace sidesteps this, processing relationships in place.

“Pathway shows that model architecture, not sheer scale, can drive the next leap in AI reasoning,” said Łukasz Kaiser, co‑author of the seminal transformer paper.

Independent verification came from NYU’s Richard Zhong and other AI veterans, confirming the ARC‑AGI‑1 scores. The team plans to expand the architecture to 600 billion parameters and target tougher benchmarks such as ARC‑AGI‑2 and ARC‑AGI‑3.

Potential applications span Reuters‑reported cybersecurity incident response and industrial optimization, where rapid, low‑cost reasoning could reshape operational economics.

Reported by Dr. Aris Thorne (Artificial Intelligence Researcher).