Pathway has published benchmark results for BDH-CQ, a 150-million-parameter reasoning model that the AI lab says establishes a new cost-efficiency point on the ARC-AGI-1 benchmark. The model achieved a 29.5% pass@2 score on the public evaluation set at a computed inference cost of $0.0007 per task, placing its performance below leading frontier systems but at a far lower estimated cost. Pathway argues that the result shows model architecture can materially change the economics of machine reasoning, particularly where repeated inference costs and latency are major constraints.
A Different Approach to Reasoning
BDH-CQ is built on Pathway’s BDH architecture, which is designed to perform reasoning through a recurrent latent state rather than by generating long chains of intermediate text. In conventional transformer-based reasoning systems, additional reasoning steps often create more tokens that are fed back into the model, increasing context length and inference expense as the process continues. Pathway says its approach instead keeps the reasoning process inside the model, allowing it to refine solutions without relying on a growing textual scratchpad.
Cost Efficiency Becomes the Main Claim
The company compared BDH-CQ with GPT-5.6 Luna at its Low reasoning setting, which the ARC Prize results page lists at 34.2% on ARC-AGI-1. Pathway said BDH-CQ delivered its 29.5% score at roughly one-eleventh of the per-task cost, even after incorporating an OpenAI price reduction announced in late July, framing the tradeoff as a modest performance gap for a substantially lower inference bill. The ARC Prize leaderboard emphasizes cost per task alongside performance, reflecting growing attention to how efficiently reasoning systems convert compute into correct answers.
External Reproduction and Infrastructure
Pathway said the results were independently reproduced by Łukasz Kaiser, a co-author of the 2017 Transformer paper, and NYU researcher Richard Zhong, who works on model evaluation and benchmark robustness. The company also trained BDH-CQ using Amazon SageMaker HyperPod, while AWS executive Nicolas Tarducci described the work as a promising route toward deploying advanced reasoning systems more economically at scale. These references do not make the benchmark result a broad measure of general intelligence, but they add validation that the reported setup and score can be reproduced outside Pathway’s internal testing.
Why ARC-AGI-1 Matters
ARC-AGI-1 tests whether a system can infer a transformation rule from a small set of examples and apply that rule to a new input, making it a closely watched benchmark for abstraction and generalization. Its tasks differ from standard language-model evaluations because success depends less on recalling information and more on identifying unfamiliar structural patterns. Pathway’s result is notable not because BDH-CQ leads in absolute accuracy, but because a relatively small model reached a competitive reasoning score while using extremely low estimated inference cost.
Pathway’s Next Steps
Pathway plans to scale the BDH architecture and test it on harder reasoning settings, including mathematical reasoning, ARC-AGI-2 and ARC-AGI-3, while also developing a larger language model that uses latent reasoning. The company says early experiments indicate that Transformer-like scaling behavior appears during pretraining from models in the billion-parameter range upward while retaining the latent-state reasoning properties demonstrated by BDH-CQ. If those characteristics hold as the architecture expands, Pathway sees potential applications in areas requiring continuous reasoning under changing information, including cybersecurity response and real-time industrial operations.
The BDH-CQ announcement shifts attention from raw benchmark leadership toward the relationship between reasoning quality, model size and inference economics. With a 29.5% ARC-AGI-1 pass@2 score at a reported $0.0007 per task, Pathway is positioning its post-Transformer architecture as evidence that smaller models can remain useful when efficiency is treated as a first-class design goal. The more consequential test will be whether the same latent-reasoning approach can preserve its cost advantage as Pathway scales the architecture and moves to harder benchmarks and production workloads.