BDH-CQ Reasoning Model Solves Puzzles Without Words at 1/11 the Cost

A small model that thinks without talking

Most reasoning AI works by writing out its thinking. Before it answers, the model produces a long chain of intermediate text, and every one of those words is a token someone pays for. A small experimental system called BDH-CQ, built by the AI company Pathway, tests a different idea: the model works through problems internally and never turns its intermediate steps into language. The researchers described the work in a paper submitted to arXiv on August 10.2 Pathway CEO Zuzanna Stamirowska, a complexity scientist, says that once a query comes in, nothing between the question and the answer is ever written out as words.2

The headline result is about cost more than raw ability. On the public ARC-AGI-1 evaluation set, a benchmark that asks a system to infer visual rules from a few examples and apply them to new puzzles, BDH-CQ solved nearly 30% of puzzles when allowed two attempts.215 The team puts the cost of each puzzle query at about $0.0007, roughly one-eleventh of what OpenAI's GPT-5.6 Luna costs on the same test.2 Live Science reports that GPT-5.6 Luna (Low) scored slightly higher, and the study calls that edge a "modest accuracy gain" that cost about 11 times as much in relative token costs.15

The framing differs by outlet. Live Science ran the claim that the model signals the start of a "post-transformer" era, and it notes that many other models score much higher on ARC-AGI-1.15 Science News gave more space to outside researchers, and it points out that the two cost figures were calculated in different ways.2 Read side by side, the two reports point to one conclusion. This is a real efficiency result. It is not yet proof that the architecture is better.

What the outside researchers say

The independent comments are useful because they separate what the work shows from what it suggests. Yuntian Deng, a computer scientist at the University of Waterloo, called it an interesting efficiency result. He also said it does not show that BDH-CQ's underlying architecture beats other approaches, and that more testing is needed to tell the effect of the design apart from the effect of the training.2 Jonas Geiping of the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems called the approach "neat" and noted that it can take on test problems without retraining. He also pointed out that BDH-CQ was built specifically for ARC-style problems, so comparing it directly with general-purpose systems is difficult.2

That second point matters most. A model tuned for one class of visual puzzles is being measured against a general model that also handles code, documents and conversation. Some of the elevenfold gap probably comes from specialization, not from a better way of thinking. The researchers argue the results suggest the model's abilities could grow a lot at larger sizes, but that is a projection, not something they have shown.15

Deng also raised a trade-off that applies to the whole approach. A model that never writes out its reasoning is harder to inspect. He added that a written chain of thought may not faithfully show how a model reached its answer anyway.2 Even so, visible reasoning traces have become a key tool for debugging and safety review. Giving them up for cheaper inference is a real cost, even if those traces were never fully reliable.

Why the cost of reasoning is the real story

BDH-CQ is getting attention because reasoning has become the most expensive part of modern AI. Reasoning models need far more computing power to answer than models that respond directly. On the AIME math competition, they have cost 10 to 74 times more than their non-reasoning counterparts.8 Those costs have pushed the industry to look for savings in several places at once.

The big labs mostly attack the problem by making their reasoning shorter. Anthropic says its Claude Opus 5.5 costs less per token than Opus 5 and also uses fewer tokens per task, so typical token-billed work costs about 40% less.4 Customer quotes on Anthropic's own page, which are not independent tests, describe the model finishing tasks in fewer steps and with far fewer output tokens than its predecessor.4 OpenAI's GPT-6 Sol and GPT-6 Luna focus on efficiency in a similar way, with prices up to 50% below the promotional pricing of their GPT-5.6 counterparts.9 Reflection AI says its open-weight Beam model, a mixture-of-experts design trained with high-compute reinforcement learning, matches leading Chinese open models on reasoning benchmarks while using three to four times less inference compute. Those claims have not been independently verified.36

All of these are gains inside the standard approach: the model still reasons by producing text, just less of it. BDH-CQ goes further and asks whether the text needs to exist at all.

Latent reasoning is becoming a research trend

Pathway is not the only group moving in this direction. A separate arXiv preprint submitted on October 13 describes HybridAgents, a framework whose listed authors include Jiaru Zou of Stanford and Together AI. It lets an AI agent switch between reasoning internally and producing readable text, and saves the text mode for tool calls and final answers.10 Tested on Qwen3-Base models with 4 billion to 14 billion parameters across 11 benchmarks, the framework reportedly used 35.5% to 54.7% fewer tokens, ran 1.8 to 2.6 times faster, and cut compute cost by roughly 63% to 73%, while averaging 10.4% better performance than the agents it was compared against.10

HybridAgents also tries to answer the inspection problem Deng raised. Its internal reasoning can reportedly be displayed as readable pseudo-text.10 The same caveats apply as with BDH-CQ: the paper has not necessarily been peer reviewed, and no one knows yet whether the approach works on much larger models.10

Skipping the reasoning step entirely is showing up in commercial products too. Cloudflare's new Clairic extraction model belongs to what the industry has started calling "System One" models, named after Daniel Kahneman's fast, intuitive mode of thinking. These models answer in a single forward pass with no reasoning and no generated output tokens, and they charge only for input.5 In a pipeline test by Cloudflare, routing extraction and decision steps through these models instead of an LLM was about 3.2 times faster and cost 49% less.5 The independent comparison is more mixed: some decision models are clearly cheaper than a GPT-6 Luna call, while others roughly tie it or cost more.17

Further out, Tufts University researchers reported a neuro-symbolic robotics system that combines neural networks with rule-based reasoning. Training it took 1% of the energy of a standard vision-language-action system, and running it took 5%. It reached a 95% success rate compared with 34% for conventional models on a structured puzzle task.7 That is a different field, but the idea is the same: brute-force generation is not the only way to get a machine to reason.

The verdict: a signal, not a revolution

The evidence supports a measured view. The "post-transformer" label attached to BDH-CQ goes further than the results do.15 A specialized model solving under a third of ARC-AGI-1 puzzles does not end the transformer. The narrower claim holds up well, though: reasoning does not always have to be spelled out, and leaving it unspoken can cut costs dramatically on the right problems.

The bigger takeaway comes from the pattern. In just a few weeks, an academic-style startup model, an open research framework and a commercial product line all bet that the expensive part of reasoning, the written-out chain of thought, is often unnecessary.2105 The big labs are getting a similar benefit within their current designs by training models to use fewer tokens.49

Two questions remain open. The first is whether latent reasoning keeps its cost advantage at the scale and generality of frontier models, which no source here shows. The second is whether the industry can accept reasoning it cannot read. HybridAgents' readable pseudo-text is one possible compromise.10 For now, BDH-CQ is best seen as an early, credible result in a shift that is already happening: as reasoning gets cheaper, models will likely do more of their thinking without showing it.