Pathway Claims 11x Lower AI Reasoning Costs With 150M-Parameter Model

Pathway’s 150-million-parameter reasoning model BDH-CQ has scored 29.5% on the public ARC-AGI-1 benchmark at a reported inference cost of $0.0007 per task, positioning the model as a significantly more cost-efficient alternative to larger reasoning models.

According to Pathway, BDH-CQ achieved a 29.5% pass@2 score on the ARC-AGI-1 public evaluation set at a computed inference cost of $0.0007 per task. The company said the model runs approximately 11 times cheaper per task than GPT 5.6 Luna (Low), which scored 34.2% on the same benchmark. Pathway described the difference as a trade-off between a modest accuracy gap and a substantial reduction in inference cost.

The company has reached a $500 million valuation, following additional funding, taking total seed funding to $30 million, and appointed Adam Kurzrok, former Google DeepMind Gemini product lead, as the CPO.

ARC-AGI-1 is designed to test whether an AI system can identify an underlying rule from a small number of examples and apply that rule to a new input. The benchmark is often used to evaluate reasoning capabilities that require models to generalise beyond patterns seen during training.

Pathway attributes BDH-CQ’s cost advantage to its architecture. While many transformer-based reasoning models generate intermediate chains of text as part of their reasoning process, BDH-CQ performs its reasoning in a recurrent latent state without producing an intermediate text trace.

The company said this allows the model to refine solutions internally rather than generating and processing additional reasoning tokens, reducing both compute requirements and latency.

“Today’s AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence. Currently, every reasoning step consumes context, adds latency, and burns compute,” said Zuzanna Stamirowska, CEO and Co-Founder of Pathway.

“We show that a different architecture changes the game and opens up a whole new space in terms of how much intelligence per dollar. A 150M-parameter model, built on Pathway’s BDH architecture, reasons recurrently in latent space and sets a new state of the art in cost efficiency on ARC-AGI-1.” 

The results were independently evaluated by Łukasz Kaiser, co-author of the 2017 ‘Attention Is All You Need’ paper that introduced the Transformer architecture. Kaiser said the results point to architecture as an important factor in advancing AI reasoning, rather than relying solely on increasing model size.

“I’ve followed Pathway closely and replicated their ARC-AGI-1 results myself. Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning,” Kaiser said.

The results were also reproduced by Richard Zhong, an NYU researcher focused on model evaluation and benchmark robustness.

Pathway said the scores and comparison costs were drawn from the ARC Prize Foundation’s public leaderboard, which compares benchmark performance with cost per task. The leaderboard data used by the company was based on submissions available as of July.

Pathway’s cost figure was calculated using measured hardware time, while the costs of comparison models reported on the leaderboard may reflect API pricing for general-purpose models.

The company plans to scale the BDH architecture and test it on more demanding reasoning benchmarks, including mathematical reasoning, ARC-AGI-2 and ARC-AGI-3. It also plans to develop a latent-reasoning large language model based on the architecture.

Pathway said early experiments indicate that Transformer-like scaling laws continue to apply during pretraining at scales ranging from 1 billion to 600 billion parameters while retaining the latent reasoning capabilities demonstrated by BDH-CQ.

The company expects the approach could eventually be applied to workloads that require models to reason continuously as information and constraints change, including cybersecurity incident response and real-time industrial operations.