A small model solved nearly three in ten ARC-AGI-1 puzzles without writing out its steps

The design point is what happens to the examples. ARC-AGI-1 hands a model a few demonstration pairs, asks it to infer the visual rule, then apply that rule to a new puzzle ^[5]. BDH-CQ uses each example to update a fixed-size memory, and that memory stays the same size as more examples arrive, so the model does not keep going back through everything it has already seen ^[4].

Zuzanna Stamirowska, a complexity scientist and CEO of the AI company Pathway, is one of the authors ^[9]. "Once the query arrives, the model doesn't write out its thinking in words at all," she said ^[7]. "Nothing in between ever converts into language," she said ^[8].

The team puts one puzzle query at about $0.00070 and calls that roughly one-eleventh of what GPT-5.6 Luna costs on the same benchmark, while noting that the two costs were calculated differently ^[10]. Multiply back and the comparison implies about $0.0077 a query for Luna ^[11]. Since the two figures came from different methods, the eleven-fold gap is an estimate of relative cost, not the same measurement made twice.

Nearly three in 10 solved with two attempts also means roughly seven in 10 puzzles went unsolved ^[12]. The misses cluster. Colour changes and certain combinations of rules gave the model more trouble than tasks that turned or moved a shape ^[13]. On harder ordering and nesting puzzles, it did better after it had seen an example of similar difficulty ^[14].

Yuntian Deng, a computer scientist at the University of Waterloo, called the finding "an interesting efficiency result" and said it does not show that BDH-CQ's underlying architecture is better than other approaches ^[15]. Deng said specialized small models like BDH-CQ are a promising route to efficient reasoning, adding that more testing is needed to separate the effects of the model's design from how it was trained ^[16]. That separation is the whole question behind the cheap per-query figure, and one benchmark run with two attempts does not perform it.

Jonas Geiping, a machine learning researcher at the ELLIS Institute Tuebingen and the Max Planck Institute for Intelligent Systems, said BDH-CQ was built specifically for ARC-style problems, which makes it hard to compare directly with general-purpose systems ^[18]. He called the approach "neat" and noted that it can take on test problems without retraining ^[18].

Deng also flagged the inspection cost. Without written-out reasoning the model is harder to examine, he said, though he added that even a written chain of thought may not faithfully show how a model reached its answer ^[17]. "Human language is useful for communicating reasoning, but it need not be the most efficient representation for every intermediate computation," Deng said ^[19].