Researchers at AI company Pathway have unveiled BDH-CQ, an experimental AI model that uses a different approach to reasoning rather than relying solely on conventional transformer architecture. Early tests suggest it can solve visual logic problems at a fraction of the cost of larger systems, potentially widening access to AI applications worldwide.
Key takeaways
- BDH-CQ scored almost 30% on the ARC-AGI-1 reasoning benchmark.
- The model uses 150 million parameters, far fewer than leading frontier systems.
- Researchers say it can cost up to 11 times less to run than a comparable OpenAI reasoning model.
- Its “latent reasoning engine” works through internal numerical states rather than long written chains of thought.
- Pathway plans to scale the architecture and test it on more difficult benchmarks.
A different route to AI reasoning
Most major systems behind products such as ChatGPT and Claude use transformers. These models process relationships between words and generate answers token by token, a method that has powered rapid progress but can become increasingly expensive as prompts and reasoning chains grow longer.
BDH-CQ takes what Pathway describes as a “post-transformer” approach. Instead of retaining a growing text-based record of its reasoning, it uses numerical arrays to represent relationships and patterns. The model then revisits its working state through repeated internal loops before producing an answer.
For Power AI Media readers tracking the next generation of AI tools, the distinction matters: the model is not simply a smaller chatbot, but an attempt to rethink how machine reasoning is stored and performed.
Lower costs, modest benchmark results
In tests using ARC-AGI-1, a benchmark built around non-verbal puzzles, BDH-CQ solved the equivalent of nearly three out of 10 tasks in two or fewer attempts. That score is below several larger systems, but the researchers argue that its efficiency is more significant than the headline accuracy alone.
The study compared BDH-CQ with OpenAI’s GPT 5.6 Luna (Low), which achieved a slightly higher result but reportedly required about 11 times the relative token cost. Tokens are the fragments of data used to measure AI input and output, and they help determine how expensive a system is to operate.
The findings remain early. The research paper was published on the preprint server arXiv and has not yet gone through conventional peer review, although the reported benchmark results were independently reproduced by researchers including NYU’s Richard Zhong and AI researcher Łukasz Kaiser.
Why the architecture could scale
BDH-CQ was trained with about 150 million parameters, compared with the tens or hundreds of billions used by many advanced models. Smaller models are generally quicker and cheaper to train, deploy and run, making them attractive for cybersecurity, industrial systems, healthcare tools and other applications where computing resources are limited.
Pathway says more reasoning loops can give the model additional time to work through a problem without increasing memory use in the same way that long text-based reasoning chains do. In theory, that could reduce GPU bottlenecks as AI systems handle more complex tasks.
What comes next
Pathway plans to expand the BDH architecture to as many as 600 billion parameters and test it on ARC-AGI-2 and ARC-AGI-3. It also intends to develop a full large language model based on the technology.
Whether BDH-CQ can match the broad capabilities, reliability and safety of established systems remains uncertain. But its results add weight to a growing argument covered across Power AI Media: future AI progress may depend not only on bigger models, but also on more efficient ways to reason.