Pathway AI Model Cuts Reasoning Costs
Researchers from AI company Pathway report their new BDH-CQ model uses a novel 'post-transformer' architecture to achieve reasoning at a fraction of the
A new artificial intelligence model from researchers at Pathway uses a novel approach to reasoning that dramatically cuts operational costs. The BDH-CQ model scored almost 30% on the ARC-AGI-1 benchmark and reportedly costs up to 11 times less to run than OpenAI's GPT 5.6 Luna (Low) model for a similar task, according to a study posted to the arXiv preprint server on August 10.
The research details a 'post-transformer' architecture. This new method processes information using numerical arrays and vectors instead of the sequential text-based reasoning of standard transformer models like those behind ChatGPT and Claude. Pathway scientists claim this allows for complex abstract reasoning without a corresponding increase in memory use or computational cost.
Performance and Cost Comparison
The 2019 ARC-AGI benchmark uses nonverbal puzzles, like rotating shapes to complete a sequence, to measure AI cognitive ability. BDH-CQ solved the equivalent of three out of ten puzzles in two or fewer attempts. While other models have achieved higher scores, the study highlights the cost efficiency of Pathway's approach.
The table below compares key figures from the source.
| Model | ARC-AGI-1 Score | Parameters | Relative Token Cost (vs. BDH-CQ) |
|---|---|---|---|
| Pathway BDH-CQ | Almost 30% | 150 million | 1x (baseline) |
| OpenAI GPT 5.6 Luna (Low) | Slightly higher | Not specified | Roughly 11x |
| Meta Llama 3 70B / 3.1 405B | Not specified | Tens to hundreds of billions | Not specified |
Pathway's model was trained on just 150 million parameters. Frontier models like Meta's Llama 3 70B use tens of billions. Fewer parameters typically mean faster training and lower running costs.
Technical Foundations of a New Architecture
Transformer models process every word in a prompt simultaneously to infer context. They generate responses word by word, creating a linear, verbal chain of thought. This method becomes computationally expensive with longer prompts, as processing power needs scale quadratically with input length.
BDH-CQ replaces this text-log system. It uses vectors, directional information stored as numerical arrays in GPU memory. These arrays track relationships between information chunks. The model employs a 'latent reasoning engine' that runs iterative loops to refine an answer, using numbers to represent problem elements. More complex problems trigger more loops, increasing time but not memory consumption.
In contrast, standard transformers achieve more thinking time by generating long text token chains. This exponentially consumes GPU memory and power across a computing cluster.
Independent Verification and Future Plans
The model's benchmark results were independently reproduced by prominent AI researchers. These include NYU researcher Richard Zhong and Łukasz Kaiser, a co-author of the seminal 2017 paper 'Attention Is All You Need' that introduced transformers. Kaiser stated he replicated the ARC-AGI-1 results and said Pathway shows model architecture, not just scale, can drive the next leap in AI reasoning.
Pathway plans to scale the BDH architecture to 600 billion parameters. The company intends to apply its vector-based reasoning to more challenging benchmarks like ARC-AGI-2 and ARC-AGI-3. They also aim to develop a full large language model for text-based chatbots. Potential applications cited include complex reasoning for cybersecurity incident response and industrial operations.
