A new artificial intelligence model developed by researchers at Pathway uses a fundamentally different approach to reasoning that dramatically cuts operational costs compared to leading AI systems, while maintaining competitive performance on cognitive benchmarks. The BDH-CQ model, detailed in a research paper published August 10 on the preprint server arXiv, achieves roughly 11 times lower token costs than OpenAI's entry-level reasoning model GPT 5.6 Luna (Low), according to the researchers.

What Makes This AI Model Different From Today's Leading Systems?

Most mainstream AI models, including those powering Claude and ChatGPT, rely on "transformer" architecture, a design that has dominated the field since 2017. Transformers work by examining every word in an input simultaneously and then predicting the next word in a response, one word at a time. While this approach has major advantages, it struggles with longer or more complex prompts because the computational complexity increases quadratically; doubling the length of an input requires four times as much processing power.

Pathway's BDH-CQ uses what the company describes as a "post-transformer" architecture that fundamentally changes how the AI stores and processes information. Instead of retaining prompts as long strings of numerical values that expand as new tokens are added, BDH-CQ uses numerical arrays to represent underlying rules and contextual patterns. This allows the model to process complex abstract reasoning without increasing its memory footprint or computational cost.

The model implements what researchers call a "latent reasoning engine" as its internal workspace. Using numbers to represent different elements of a problem, it carries out iterative loops to determine the best answer. The model takes the output of one loop, assesses how the result could be improved based on its training data, and feeds that output back as the starting point for the next iteration. Crucially, this approach means that running 200 loops consumes roughly the same amount of memory and computing power as running 20 loops, whereas standard transformer models exponentially consume GPU memory and computing power with each additional thinking step.

How Does BDH-CQ Perform on Reasoning Benchmarks?

The researchers evaluated BDH-CQ against ARC-AGI-1, a foundational 2019 benchmark designed to measure progress toward artificial general intelligence (AGI), the theoretical point at which AI matches or surpasses human capabilities across all domains. ARC-AGI uses nonverbal reasoning puzzles, such as rotating shapes to complete sequences, to test cognitive ability in ways that require inferring rules through trial and error, much like humans do.

BDH-CQ scored approximately 30 percent on the ARC-AGI-1 benchmark, successfully solving the equivalent of three out of 10 puzzles in two or fewer attempts. While numerous models have achieved significantly better scores on this test, the underlying reasoning approach makes BDH-CQ's size and usage costs dramatically smaller than models built on traditional transformer architecture.

The efficiency gains are striking when compared to larger systems. BDH-CQ was trained on just 150 million parameters, while advanced frontier AI models such as Meta's Llama 3.1 405B typically number tens of billions to hundreds of billions of parameters. In AI development, fewer parameters means models are faster to train and cheaper to run.

Why Does This Matter for AI Deployment?

The cost advantage of BDH-CQ has significant implications for how AI systems are deployed at scale. AI model usage is measured on a per-token basis, with each token representing roughly four characters of text that the AI must ingest or output. Because more complex prompts require longer trains of thought with multiple steps, processing and responding to these queries can consume substantial amounts of tokens, driving up operational costs.

Conventional transformer-based token generation is prone to memory bottlenecks, as AI re-reads every previous word in a conversation with every new word generated. Eventually, this clogs up the memory in the graphics processing units (GPUs) used for AI operations. Pathway's post-transformer approach changes how the AI's memories and information relationships are stored and processed, replacing text logs with improved short-term memory and mechanisms that allow the model to work through problems without consuming tokens.

How to Evaluate the Real-World Impact of This Research

• Cost Efficiency Gains: Compare the token costs of BDH-CQ to existing models when deployed for reasoning-heavy tasks; the 11-fold reduction in costs relative to GPT 5.6 Luna suggests significant savings for organizations running AI at scale.
• Scalability Potential: Monitor whether the model's performance improves when trained on larger parameter sizes; researchers believe cognition capabilities could scale significantly with expansion, potentially closing the accuracy gap with larger models.
• Benchmark Reproducibility: Verify that independent researchers can reproduce BDH-CQ's ARC-AGI-1 results; the study notes that prominent researchers including NYU's Richard Zhong and Łukasz Kaiser, co-author of the seminal 2017 transformer paper, have independently verified the findings.

The researchers believe that if this post-transformer architecture is adopted widely, it could have a dramatic impact on the overall cost and scale of AI deployments. The model's approach to reasoning through iterative refinement rather than sequential token generation represents a meaningful departure from how leading AI systems currently operate.

BDH-CQ builds on earlier work by Pathway scientists, who created a precursor model known as "Dragon Hatchling" in 2025, designed to accurately simulate how neurons in the brain connect and strengthen during learning. The new model represents an evolution of that approach, applying those principles to a more efficient reasoning framework.

The independent verification of BDH-CQ's results by prominent researchers in the AI field adds credibility to the findings. As AI systems become increasingly expensive to operate, alternative architectures that maintain competitive reasoning performance at a fraction of the cost could reshape how organizations approach AI deployment and infrastructure planning.