Pathway research team claims new ARC-AGI-1 cost-efficiency record with 150M-parameter BDH-CQ model

A small research group called Pathway has published a paper claiming a new state of the art in cost-efficiency on ARC-AGI-1, one of the benchmarks used to measure abstract reasoning in AI systems, using a compact 150-million-parameter model that costs a fraction of a cent per solved task.

What's new

The paper, "BDH-CQ: In-Context Learning with Recurrent Latent Reasoning," was submitted to arXiv on August 10, 2026 by a team including Björn Engdahl, Adrian Kosowski, Jan Chorowski, Zuzanna Stamirowska, and other Pathway-affiliated researchers. The authors report their model reaches 29.5% pass@2 on ARC-AGI-1 at roughly $0.0007 per task — under one-tenth of a cent — which they say beats the previously reported accuracy-per-dollar tradeoff on that benchmark. In the paper's own words: "This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency."

BDH-CQ combines in-context learning with what the authors call recurrent latent reasoning, built on the "BDH" (Dragon Hatchling) architecture Pathway has been developing as an alternative to standard transformer designs. Rather than relying on a large parameter count, the model leans on iterative reasoning passes at inference time to extract more capability per parameter. The authors report the results include independent black-box verification by evaluators outside the core team, and the paper discusses failure modes and limitations alongside its headline results rather than presenting only favorable numbers.

By the authors' accounting, the cost-efficiency gain works out to roughly 57 times cheaper than running the same class of task on OpenAI's GPT-5.6 Luna at list price, or about 11 times cheaper once OpenAI's mid-2026 price cut to Luna is factored in.

Context

ARC-AGI-1 has become a closely watched benchmark precisely because it's difficult to game with scale alone — it's designed to test novel abstract-reasoning generalization rather than pattern memorization, and frontier labs including OpenAI have highlighted their own ARC-AGI scores as evidence of reasoning progress. Most of the public attention on ARC-AGI performance has gone to costly frontier models running extensive test-time compute. BDH-CQ approaches the same benchmark from the opposite direction: a small model, cheap per-task inference, and an architecture built around recurrence rather than raw parameter scale.

Pathway is not one of the major labs tracked for frontier chat products — it's a smaller research outfit that has published prior work on the BDH architecture as a departure from mainstream transformer scaling. This paper continues that thread rather than announcing a product.

Why it matters

If the cost-accuracy tradeoff holds up under broader scrutiny, it's a data point against the assumption that better reasoning benchmarks require ever-larger, ever-more-expensive models. A 150M-parameter model beating the cost-efficiency frontier on a reasoning benchmark — even at a modest 29.5% pass rate — suggests architectural choices around recurrence and in-context learning can substitute for scale on at least some classes of reasoning tasks. That matters most for cost-sensitive deployment: applications running millions of reasoning queries per day care as much about dollars per task as raw accuracy, and a model this cheap changes what's economically viable to run at scale. It's a narrow result on one benchmark from a smaller lab, not a general claim of parity with frontier models, and the modest absolute pass rate (29.5%) means it's not competing with top frontier scores on raw capability — but the independent verification and honest limitations discussion give the result more credibility than a typical unverified benchmark claim.