BDH-CQ: A 150M-Parameter Model Scored 29.5% on ARC-AGI-1 at $0.0007 per Task

Pathway researchers reported in "BDH-CQ: In-Context Learning with Recurrent Latent Reasoning," posted to arXiv on August 10, 2026, that a 150M-parameter configuration reaches 29.5% pass@2 on the public ARC-AGI-1 evaluation set at a computed inference cost of $0.0007 per task. The authors state that this operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier and establishes a new state of the art in benchmark cost efficiency. The paper lists nine authors, including Adrian Kosowski, Jan Chorowski, and Richard Zhong.

Reasoning Happens in Latent Space and Is Never Verbalized

The defining choice in BDH-CQ is that intermediate reasoning runs as iterative computation in a high-dimensional continuous latent space and only the final answer is decoded. Demonstration inputs presented at inference time continuously update the model's recurrent memory, and the query is then solved by repeated computation over that memory. Because no chain-of-thought text is generated, adding reasoning steps does not add output tokens.

The paper separates two states with different jobs. Contextual memory S_t changes as evidence arrives and supports in-context learning, while reasoning workspace H_r carries the ongoing computation used to answer the current query. That separation is the architectural claim. Keeping what was learned apart from what is being computed means extracting a rule from demonstrations and applying it to a new input need not share one token stream.

The foundation is the Dragon Hatchling (BDH) line Pathway published in 2025, described here as ReLU-low-rank transformations with linear attention in a large neuron or feature space. Unlike a transformer attention cache that grows with context length, running more recurrent iterations does not expand memory the same way, and that property is the root of the cost argument.

The Cost Advantage Sits Between 57x and 11x, Pinned to Someone Else's Price List

The reported cost advantage moves between 57x and 11x, and the pivot is which date's public API price is used for the comparison model, GPT-5.6 Luna (Low). Using ARC Prize's reported costs as of July 2026, where GPT-5.6 Luna (Low) scores 34.2% at $0.040 per task, the authors put BDH-CQ at roughly 57x cheaper. Accounting for OpenAI's 80% public API price reduction of that model on July 30, 2026, the same paragraph states the figure narrows to roughly 11x.

Two numbers coexisting in one paper defines what kind of result this is. In the Pareto claim, the vertical axis is a property of the model while the horizontal axis is a competitor's pricing policy. One more OpenAI price cut shrinks the multiple again, and a price increase widens it, all without a single change to the architecture.

The accuracy gap, unlike the price, does not move. GPT-5.6 Luna (Low) scores 34.2% on the same evaluation against 29.5% for BDH-CQ, a 4.7-point difference that a cost-multiple headline tends to hide. Stated precisely, the result is not "solves more" but "lets you attempt far more for the same money."

$0.0007 Is Converted GPU-Seconds, Not an API Invoice

The $0.0007 per task figure is a computed cost from 0.85 H200 GPU-seconds priced at $3 per H200-hour, not a commercial API rate. The paper states that at the default operating point the 150M-parameter system reaches 29.5% pass@2 in approximately 0.85 H200 GPU-seconds per task, and multiplies through to $0.00070.

The comparison number lives on a different accounting layer. The $0.040 for GPT-5.6 Luna (Low) is a sale price carrying OpenAI's margin and operating overhead, while $0.0007 is a self-run unit cost. Putting them on one basis would require adding serving overhead, utilization, and batching efficiency to the BDH-CQ side, and OpenAI's own cost is not knowable from outside. That asymmetry is why the authors label theirs a computed cost.

The method still has practical value. Teams running models on their own hardware budget in GPU-seconds per task, not in API line items, and 0.85 seconds is a verifiable unit cost. Paired with the 150M parameter count, it lets an engineer estimate per-GPU throughput directly.

The Failure Map From the Interventions Carries More Information Than the Score

The controlled ARC-like interventions show, item by item, which generalizations hold and which collapse. Rule propagation extrapolates fully at 48/48 held-out outputs at both pass@1 and pass@2 across distances 2 through 8, and copying holds at 48/48 as target sites increase from one to four. The ability to extract a transformation from demonstrations and apply it repeatedly is solid.

Three things break. Ordering is nearly saturated through length 5, then falls to 29/36 at length 6, 8/24 at length 7, and 1/24 at length 8. Nesting is nearly saturated through depth 4 and drops to 29/36 at depth 5. Support-chain length declines monotonically at 80.0% with one object, 67.5% with three, 52.5% with five, and 27.5% with eight.

Composition is sharper still. Rotation composed with relocation succeeds at 72/72, while reflection composed with relocation reaches only 47/72. Color swap scores 26/72 on the original motif family alone and 0/72 once composed with relocation. Conditional rule selection, where corner markers decide which of two rules applies, lands at 56.7%.

One sentence summarizes the map. BDH-CQ is strong at extracting a single rule and pushing it far, and weak at binding two rules together or holding several states at once. Nothing inside the 29.5% headline reveals what the latent-iteration design buys and sells as clearly as this section does, which makes the intervention protocol the most reusable part of the paper.

The Training Mix Limits Any Reading of This as a General Reasoner

BDH-CQ trained on a curated collection of ARC-style data combining privately curated examples with public sources including ARC-AGI-1, RE-ARC, ConceptARC, ARC-Heavy, and ARC-GEN100K, plus augmentations. That list, disclosed in the paper, bounds the claim. This is a specialist trained for ARC-format grid transformation, and it asserts nothing about language understanding or code generation.

Being a specialist does not diminish the result; it is closer to the actual thesis. For narrowly defined structural reasoning, 150M parameters and under a second of GPU time reach the 30% band where frontier models operate. The summary "cut reasoning cost 11x," however, becomes a claim the paper never made the moment it is read as substitutability for a general model.

One more caveat belongs here. A model trained on data from the same family as the public evaluation set and then scored on that public set does not give a number identical to expected performance on a fully unseen task distribution. The composition failures in the intervention section mark exactly where that limit shows up in practice.

A State-of-the-Art Result Published in an Unreproducible Form

BDH-CQ is a state-of-the-art cost-efficiency result that no outside team can reproduce today, because the paper states that dimensions, exact update rules, and implementation details remain proprietary and contains no mention of releasing weights or training code. The number of latent reasoning iterations R appears only in general equations, with no evaluation-time value disclosed. Training compute budget and duration are also undisclosed.

Under those conditions, outsiders can check consistency rather than reproduce results. Answers on the public ARC-AGI-1 evaluation set are checkable by anyone, so 29.5% becomes verifiable the moment weights ship. The 0.85 GPU-seconds figure, by contrast, has no verification path outside the authors' report until the implementation is public.

That also fixes the weight of the Pareto claim. Moving a frontier is verifiable only on a reproducible coordinate system, and here the vertical axis of 29.5% is confirmable while the horizontal axis of $0.0007 rests on the authors' hardware assumptions. Half the Pareto claim is externally checkable.

What Practitioners Should Actually Take From This

The transferable piece is not the proprietary architecture but the practice of treating GPU-seconds per task as the cost metric. Teams that select models purely on API price versus accuracy have no coordinate for computing the break-even point of moving to self-hosted serving. Measuring GPU-seconds per task first turns the choice between a frontier API and a small specialist into arithmetic rather than preference.

The second piece is the intervention format. Splitting capability into axes such as propagation, copying, ordering, nesting, composition, and conditional selection, then measuring where each one collapses, transfers to any domain. For document processing or table transformation alike, a model whose accuracy bends at length 6 will fail in deployment on length-10 data.

What is established as of August 2026 is that on narrow structural reasoning, 150M parameters reached 29.5% pass@2 in under a second of GPU time. Whether that extends beyond ARC is outside what the paper measured, and it stays unanswerable from outside until the proprietary implementation is released.

Source: arXiv paper "BDH-CQ: In-Context Learning with Recurrent Latent Reasoning" (arXiv:2608.09888, submitted August 10, 2026; authors Björn Engdahl, Adrian Kosowski, Jan Chorowski, Zuzanna Stamirowska, Przemysław Uznański, Junlin Jiang, Rohan Phadke, Remigiusz Kinas, Richard Zhong. 150M-parameter configuration at 29.5% pass@2 on the public ARC-AGI-1 evaluation set, approximately 0.85 H200 GPU-seconds per task, $0.00070 computed at $3 per H200-hour; roughly 57x cheaper than GPT-5.6 Luna (Low) at 34.2% and $0.040 per ARC Prize costs as of July 2026, roughly 11x after OpenAI's 80% price reduction on July 30, 2026; contextual memory S_t separated from reasoning workspace H_r, ReLU-low-rank transformations with linear attention; training data spanning ARC-AGI-1, RE-ARC, ConceptARC, ARC-Heavy, ARC-GEN100K and private curation; interventions covering propagation 48/48, copying 48/48, ordering 29/36 at length 6, 8/24 at length 7, 1/24 at length 8, nesting 29/36 at depth 5, rotation composition 72/72, reflection composition 47/72, color swap 26/72 alone and 0/72 composed, conditional selection 56.7%, support chain from 80.0% to 27.5%; dimensions, update rules and implementation details stated as proprietary) compiled by ASAP.