News · 2026-08-11

A 150M model set an ARC-AGI record for cost, not score

Pathway published BDH-CQ, a 150-million-parameter model that solves abstract reasoning puzzles without writing out its thinking, and it set a record on the public ARC-AGI-1 benchmark -- for cost, not for score. The model reaches 29.5% pass@2 at a computed inference cost of seven hundredths of a cent per task, which the authors describe as a new state of the art in benchmark cost efficiency. It became the most upvoted paper of the day on Hugging Face.

Key facts

• A 150-million-parameter configuration reaches 29.5% pass@2 on public ARC-AGI-1 at a computed $0.0007 per task.
• The claim is explicitly a cost-efficiency record, not an accuracy record.
• #1 paper of the day on Hugging Face with 247 upvotes, submitted 10 August 2026 by a Pathway team including Bjorn Engdahl, Adrian Kosowski, Jan Chorowski and Zuzanna Stamirowska.
• Primary source: BDH-CQ: In-Context Learning with Recurrent Latent Reasoning, with a company blog post and a task-generation repo.

Thinking without talking

ARC-AGI is a benchmark of visual puzzles. Each task shows a handful of before-and-after grids that demonstrate some transformation, then gives you a new grid and asks what the answer should be. There is no memorising the answer key; the rule is different every time. It is deliberately built to test whether a system can infer a novel rule from a few examples -- which is why it has been a stubborn benchmark for models that are otherwise superb at everything else.

The dominant approach for the last two years has been to make the model think out loud. Chain-of-thought reasoning works, and modern systems push it further by spending more inference time on harder problems -- what the field calls test-time compute. It also costs money in direct proportion to how much the model says, because every step of reasoning is generated text.

BDH-CQ does something else. In the authors' description, "inputs presented at inference time continuously update the model's recurrent memory; the model then solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning."

Two separate stores, two separate jobs. The recurrent memory absorbs the demonstrations -- this is in-context learning implemented as state rather than as a long prompt. The latent workspace carries the ongoing computation, iterating on the problem internally. Nothing gets written down.

The analogy that fits is the difference between doing arithmetic on paper and doing it in your head. Working on paper is legible and checkable, and every line costs a pen stroke. Doing it in your head is opaque, but the steps are effectively free. BDH-CQ does its arithmetic in its head, and the price of a task drops accordingly. Related: looped transformers and weight tying.

Read the claim precisely

Here is where the reporting around this result needs care. The paper says the 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."

Cost efficiency. Not accuracy. The frontier reasoning systems score far higher on the same benchmark and cost orders of magnitude more per task. What Pathway has demonstrated is a new point on the curve -- radically cheaper per point of accuracy -- not a new peak. Anyone reading this as a small model beating large ones on capability has misread it, and the paper does not encourage that reading.

For context on why this distinction matters constantly, see how AI is benchmarked and our earlier coverage of how two API settings tripled OpenAI's ARC-AGI-3 score.

Why it matters anyway

The cost figure is the interesting part precisely because it is so far from the frontier's economics. Seven hundredths of a cent per task, from a model small enough to run on hardware most developers already own, on a benchmark specifically designed to require novel rule inference. That is an existence proof that at least some reasoning work is bottlenecked on computational structure rather than on scale.

If iterative computation in a latent workspace can substitute for verbalised reasoning at a fraction of the cost, the implications reach well past one puzzle benchmark -- into every deployment where a model has to think about many items and the thinking is the bill. Whether it generalises is the open question, and it is the one the paper does not answer.

The honest caveat

Two limits. First, ARC Prize's community leaderboard treats submitted scores as self-reported unless separately verified, so this is a claim from the authors rather than a certified placement. Second, and more substantively: a 150M model being very cheap per point on one benchmark tells you almost nothing about whether latent reasoning holds up on open-ended tasks, long-horizon work, or anything requiring the model to explain itself. Reasoning you cannot read is reasoning you cannot audit, which is a real cost that does not appear in the dollar figure.

Primary source, verified: read the paper → (arXiv 2608.09888)

Ground Truth. (2026, August 11). A 150M model set an ARC-AGI record for cost, not score. Ground Truth. https://groundtruth.day/news/a-150m-model-set-an-arc-agi-record-for-cost-not-score.html

@misc{groundtruth:a-150m-model-set-an-arc-agi-record-for-cost-not-score,
  title  = {A 150M model set an ARC-AGI record for cost, not score},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/a-150m-model-set-an-arc-agi-record-for-cost-not-score.html}
}

Topics: reasoning · arc-agi · architecture · efficiency · latent-reasoning · small-models