Pathway’s BDH-CQ, a 150M-parameter post-transformer reasoning model, scored 29.5% pass@2 on the public ARC-AGI-1 evaluation set at a computed inference cost of $0.0007 per task.
That is roughly 11 times cheaper per task than GPT 5.6 Luna (Low), even after OpenAI cut Luna’s price by 80% on July 30. If the number survives outside the press release, it rewrites the cost-per-task math that decides which reasoning workloads are actually deployable.
Here is what changed, what is still an assumption, and what I would do about it.
What the benchmark numbers actually say
BDH-CQ hit 29.5% pass@2 across the 400-task public ARC-AGI-1 evaluation set, Pathway’s first public results for the model.
GPT 5.6 Luna (Low) scored 34.5% on the same benchmark.
So this is not a raw-accuracy win.
BDH-CQ trails Luna by about five points on accuracy while running roughly 11 times cheaper per task.
The claim is about the frontier, not the top of the leaderboard.
The cost figure comes with a clear method.
Each task took about 0.85 seconds of NVIDIA H200 GPU time. And Pathway assumes an H200 cost of $3 per GPU-hour, which lands at approximately $0.00070 per task. That is a calculated cost, not a billed one. The H200 rate is an assumption baked into the math. And your actual per-task cost depends on what you actually pay for H200 capacity. Treat $0.0007 as a defensible estimate, not an invoice.
The result is framed two ways, and both are worth separating.
Pathway says BDH-CQ “has set the new cost-efficiency frontier on ARC-AGI 1.” The Hugging Face paper summary calls the same operating point a break through “the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.” Same claim, two registers. The point is that on the cost-to-accuracy curve, BDH-CQ sits in front of where any prior model sat.
Why a 150M model can compete at all
This is the part the press releases hand-wave, so let me be blunt about what little we know.
The Hugging Face summary describes BDH-CQ as using “recurrent latent reasoning and in-context learning.” That is the whole architectural story in one line, and it matters.
Most reasoning models you run today externalize their work. They generate intermediate tokens, feed those tokens back in, and repeat until they reach an answer. Every step is text you pay to generate and pay to read back in. That is the compute tax that makes reasoning models expensive. And it is why a cheap reasoning workload feels like a contradiction in terms.
Latent reasoning does the work somewhere other than the output token stream.
The reasoning happens inside a recurrent state, not as a visible chain of text the model re-reads.
From a cost standpoint, that is the load-bearing difference. If the intermediate steps do not leave the model as tokens, they do not get billed as tokens. A 150M-parameter model can plausibly compete on ARC-AGI-1 not because it is smarter but given that it is not paying the chain-of-thought tax that frontier reasoning models pay on every task.
I am overstating the cleanliness of this slightly as the public summary is thin.
We do not have the full mechanism for how the recurrent latent reasoning is implemented, how it scales, or where it breaks.
But the direction is consistent with the cost number. You do not get to 11 times cheaper on the same architecture doing the same work.
Something structural is different.
What this means if you pay for reasoning at scale
Run the builder math. If you have a workload doing thousands or millions of reasoning passes a day, per-task cost is the number that decides whether the product ships or stays a demo. A workload that cost X on Luna costs roughly X divided by 11 on BDH-CQ, accepting the five-point accuracy gap. That is not a rounding improvement. It is the difference between a feature you meter aggressively and a feature you let run.
This matters most for the workloads that were never viable at frontier pricing. Real-time adaptation, per-user personalization, any job that wants to call a reasoning model inside a tight loop instead of once per session. Those are the use cases that die at Luna’s per-task cost and survive at roughly an eleventh of it. Pathway frames this as a “sustainable path toward maximizing intelligence per dollar,” and for once that is not just marketing language. The cost-efficiency frontier moving left is exactly what unlocks the jobs you previously could not afford to ship.
I run client automations where cost-per-call is the constraint that kills half the ideas before they reach production.
A drop like this does not mean I rewrite everything tomorrow.
It means a class of workflow I rejected on cost grounds two months ago is back on the table. And I owe it a second look.
The skeptical read, since I have been burned
A few caveats before you start replanning your roadmap. The score is 29.5%, not 95%. ARC-AGI-1 is a benchmark designed to test whether a system can infer transformations from a small number of examples and apply those rules to unseen inputs. And a sub-30% score means the model fails most of the tasks. It is winning on cost-efficiency, not on capability.
The 29.5% is pass@2, which means the model gets two attempts at each task.
Single-shot accuracy is lower. The evaluation is on the public set, which is the set developers can tune against. And the cost figure depends on the H200 assumption, which is reasonable but not guaranteed to match your own deployment reality. This is also first public results for BDH-CQ, which means nobody has stress-tested the architecture at the scale or the failure modes that production surfaces.
Benchmark wins have a bad habit of not surviving contact with real workloads.
I have watched enough “10x cheaper, just as good” claims age poorly that I want a second source running the same evaluation before I trust the frontier claim.
Until then, I treat this as a strong signal that the architecture is serious, not as proof that it will hold at the scale you actually care about.
What I would actually do
Watch Pathway, and watch whether the result gets independently reproduced. The signal worth tracking is not the leaderboard score but whether the cost-efficiency frontier keeps moving left as other teams adopt or adapt the latent-reasoning approach. If a 150M model can hold a frontier position, the assumption that intelligence requires scaling token generation starts to crack.
For the practical case, if you run reasoning workloads at volume, pull the Hugging Face summary and the public ARC-AGI-1 results and sanity-check the cost math against your own GPU rates. The number that matters for your business is not $0.0007. It is what BDH-CQ, or whatever inherits its approach, costs on your infrastructure at your task mix. That is the number that tells you whether to rebuild. Everything else is a press release.
If you are building or auditing AI automations and want help modeling what an order-of-magnitude cost shift does to your unit economics, that is exactly the work my agency does.
Reach out and we will run the numbers on your actual workload.
