Pathway has introduced BDH CQ, a tiny 150 million parameter reasoning model. The new architecture processes logic within a recurrent latent state rather than generating a long chain of thought text. However, it still trails the accuracy of larger models like OpenAI GPT 5.6 Luna.
Data from the public ARC Prize Foundation leaderboard shows that the model manages to complete tasks at an estimated cost of 0.0007 dollars per task. This budget friendly footprint makes the model roughly 11 times cheaper than GPT 5.6 Luna Low. OpenAI recently slashed the price of Luna by 80% to remain competitive, yet the cost discrepancy remains massive. While Luna achieved a 34.2% accuracy score on the ARC AGI 1 reasoning test, the smaller model held steady at 29.5%.
Standard reasoning systems run up high costs because they write out intermediate steps as text tokens. As this trace grows, computing costs and latency spike. The new architecture from Pathway thinks natively inside its hidden layers without relying on an external text scratchpad. Zuzanna Stamirowska, the chief executive officer at the lab, explained that architecture has created an unnecessary bottleneck for artificial intelligence.
Today's AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence. A 150M parameter model reasons recurrently in latent space, and sets a new state of the art in cost efficiency. The bottleneck was never intelligence. It was design.
The results have already been verified by external researchers. Lukasz Kaiser, a co author of the original 2017 paper that introduced the Transformer architecture, replicated the test results and confirmed the findings. Richard Zhong, an evaluator from New York University, also reproduced the benchmarks. Early testing suggests that this new approach can scale effectively. Pathway has run experiments on models ranging from 1 billion parameters up to 600 billion parameters, finding that the architecture scales in a predictable pattern similar to traditional transformer models.
The research lab is currently training the system on Amazon SageMaker HyperPod. Nicolas Tarducci from Amazon Web Services noted that this work offers a promising way to deploy reasoning systems cost effectively at scale. Moving forward, Pathway plans to apply the same concept to more difficult domains like mathematical reasoning, ARC AGI 2, and ARC AGI 3, with the goal of building a larger language model that relies entirely on latent reasoning.
