Blab AI
Pathways 150-M-Parameter BDH-CQ Model Sets New Cost-Efficiency Record on ARC-AGI-1 Benchmark
In a move that could reshape how we evaluate and deploy reasoning AI, Pathway announced on 11 August 2026 that its 150‑million‑parameter model, BDH‑CQ, earned a 29.5 % pass@2 score on the public ARC‑AGI‑1 benchmark—doing so at an inferred cost of just $0.0007 per task.
ARC‑AGI‑1 measures a model’s ability to infer a rule from a handful of examples and apply it to a new input, a standard test for reasoning. While BDH‑CQ trails OpenAI’s GPT‑5.6 Luna by 4.7 percentage points (34.2 % pass@2), its inference cost is roughly eleven times lower. The comparison relies on hardware‑time data reported by Pathway for BDH‑CQ and on the API pricing listed on the leaderboard for GPT‑5.6 Luna.
The key to this efficiency lies in BDH‑CQ’s architecture. Unlike transformer‑based models that generate a chain‑of‑thought token sequence, BDH‑CQ reasons in a recurrent latent state, learning from demonstrations and refining its solution internally. By avoiding the extra tokens that inflate latency and compute, Pathway claims the token‑based cost that has constrained previous reasoning systems is eliminated.
"Today’s AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence," said Zuzanna Stamirowska, Pathway’s CEO and co‑founder. "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 150‑M‑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 bottleneck was never intelligence. It was designed," she added.
The benchmark results were independently replicated by Łukasz Kaiser, co‑author of the 2017 transformer paper, and by Richard Zhong, an NYU researcher focused on model evaluation. Both researchers confirmed the reported score and the cost advantage.
Alongside the announcement, Pathway disclosed a new funding round that values the company at $500 million, bringing its total seed capital to $30 million. The round includes Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital, WS Investment Co. (the investment arm of Wilson Sonsini), and angel investor Jonathan Frankle, chief AI scientist at Databricks. Capital will be directed toward expanding compute capacity, including new GB300 systems.
Pathway also named Adam Kurzrok—former group product manager for Gemini at Google DeepMind—as chief product officer. Kurzrok will steer product direction for BDH‑based models, overseeing packaging, evaluation, and large‑scale deployment.
Looking ahead, Pathway plans to scale the BDH architecture to larger models and to tackle more demanding reasoning benchmarks such as ARC‑AGI‑2 and ARC‑AGI‑3. The company also aims to develop a latent‑reasoning large language model that could support applications requiring reliable reasoning as constraints change, including cybersecurity incident response and real‑time industrial operations.
Early experiments suggest that transformer‑like scaling laws hold during pre‑training for BDH models ranging from 1 B to 600 B parameters, while preserving the latent‑reasoning capability demonstrated by BDH‑CQ.
The announcement comes at a time when AI startups are increasingly scrutinizing the cost‑efficiency of large language models. BDH‑CQ’s performance demonstrates that architectural innovations can reduce inference costs without sacrificing accuracy—a finding that may influence future research and commercial deployment of reasoning‑capable AI.
As Pathway moves into its next phase of model development and commercialization, its focus on efficient reasoning and its growing investor base position it to play a significant role in the evolving AI ecosystem.
ARC‑AGI‑1 measures a model’s ability to infer a rule from a handful of examples and apply it to a new input, a standard test for reasoning. While BDH‑CQ trails OpenAI’s GPT‑5.6 Luna by 4.7 percentage points (34.2 % pass@2), its inference cost is roughly eleven times lower. The comparison relies on hardware‑time data reported by Pathway for BDH‑CQ and on the API pricing listed on the leaderboard for GPT‑5.6 Luna.
The key to this efficiency lies in BDH‑CQ’s architecture. Unlike transformer‑based models that generate a chain‑of‑thought token sequence, BDH‑CQ reasons in a recurrent latent state, learning from demonstrations and refining its solution internally. By avoiding the extra tokens that inflate latency and compute, Pathway claims the token‑based cost that has constrained previous reasoning systems is eliminated.
"Today’s AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence," said Zuzanna Stamirowska, Pathway’s CEO and co‑founder. "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 150‑M‑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 bottleneck was never intelligence. It was designed," she added.
The benchmark results were independently replicated by Łukasz Kaiser, co‑author of the 2017 transformer paper, and by Richard Zhong, an NYU researcher focused on model evaluation. Both researchers confirmed the reported score and the cost advantage.
Alongside the announcement, Pathway disclosed a new funding round that values the company at $500 million, bringing its total seed capital to $30 million. The round includes Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital, WS Investment Co. (the investment arm of Wilson Sonsini), and angel investor Jonathan Frankle, chief AI scientist at Databricks. Capital will be directed toward expanding compute capacity, including new GB300 systems.
Pathway also named Adam Kurzrok—former group product manager for Gemini at Google DeepMind—as chief product officer. Kurzrok will steer product direction for BDH‑based models, overseeing packaging, evaluation, and large‑scale deployment.
Looking ahead, Pathway plans to scale the BDH architecture to larger models and to tackle more demanding reasoning benchmarks such as ARC‑AGI‑2 and ARC‑AGI‑3. The company also aims to develop a latent‑reasoning large language model that could support applications requiring reliable reasoning as constraints change, including cybersecurity incident response and real‑time industrial operations.
Early experiments suggest that transformer‑like scaling laws hold during pre‑training for BDH models ranging from 1 B to 600 B parameters, while preserving the latent‑reasoning capability demonstrated by BDH‑CQ.
The announcement comes at a time when AI startups are increasingly scrutinizing the cost‑efficiency of large language models. BDH‑CQ’s performance demonstrates that architectural innovations can reduce inference costs without sacrificing accuracy—a finding that may influence future research and commercial deployment of reasoning‑capable AI.
As Pathway moves into its next phase of model development and commercialization, its focus on efficient reasoning and its growing investor base position it to play a significant role in the evolving AI ecosystem.
