Pathway has released results for an artificial intelligence (AI) reasoning system called BDH-CQ that contains 150 million parameters. On the public ARC-AGI-1 evaluation set the system scored 29.5 percent when allowed two attempts. The computed cost of each task came to 0.0007 dollars, less than one tenth of a cent. By comparison, a recent version of GPT 5.6 Luna scored 34.2 percent but at roughly eleven times the cost, even after a large price reduction.
BDH (Dragon Hatchling) is a post-transformer AI architecture. ARC-AGI-1 is a public test designed to measure whether a system can discover an underlying rule from a handful of examples and then apply that rule to a new case. Success on this test is often linked to flexible, human-like reasoning. The cost advantage of BDH-CQ stems from a basic difference in how the work is performed. Most systems built on the widely used transformer design produce a chain of intermediate text steps that are fed back into the model. Each extra step increases the amount of computation required and the time needed for an answer. BDH-CQ instead carries out the same process inside a recurrent latent state, an internal updating memory that refines the solution without ever writing out intermediate words.
How the new approach differs
The architecture keeps both memory and reasoning inside the model rather than expanding an external context window. Early experiments show that the same design continues to follow familiar scaling patterns when the parameter count rises from one billion to six hundred billion, while still retaining the ability to reason in the internal state. Independent checks of the ARC-AGI-1 numbers were performed by outside researchers. Pathway intends to apply the same methods to harder mathematical tests and larger language models so that reliable reasoning can support practical tasks such as responding to cybersecurity incidents or managing changing industrial conditions. The reported results therefore point to a possible shift in which design choices, rather than sheer size alone, determine how much useful intelligence can be obtained for a given amount of computing effort.
Łukasz Kaiser, co-author of the original 2017 Transformer paper, replicated the ARC-AGI-1 results and stated that Pathway demonstrates how “model architecture, not just scale, can drive the next leap in AI reasoning.” He also posted to X: "If you think prices of tokens can't go down, you've seen nothing yet. Research still delivers strong wins :)."