Palo Alto, California-based Pathway, an AI lab developing post-Transformer models, said on August 11 that its BDH-CQ model scored 29.5% pass@2 on ARC-AGI-1 at a computed cost of US$0.0007 per task.

ARC-AGI-1 is a public benchmark that measures fluid intelligence on novel pattern tasks. Its pass@2 scoring gives models two attempts because some tasks can be ambiguous.

Pathway said OpenAI’s GPT-5.6 Luna (Low) scored 34.2% on the same test, based on ARC Prize Foundation leaderboard data from July 2026. It is said the model cost about 11x more per task.

The company said BDH-CQ avoids producing intermediate text and can reduce reasoning overhead.

Pathway said Łukasz Kaiser replicated BDH-CQ’s ARC-AGI-1 results, and that Richard Zhong, a New York University researcher and co-author from Bielik, also reproduced them.

🔗 Source: Pathway