Polish Startup Pathway Sets "Intelligence per Dollar" Record in AI

Polish company Pathway has announced that its experimental language model has set an unofficial cost-efficiency record in artificial intelligence. The BDH-CQ model, with just 150 million parameters, achieved a score on the ARC-AGI-1 benchmark comparable to one of OpenAI's GPT-5.6 variants, but at eleven times lower cost per task.

Architecture over scale

At the core of the achievement is Pathway's proprietary Dragon Hatchling (BDH) architecture, which the company is developing as an alternative to the transformer architecture that has dominated the field for years and underpins models like GPT and Gemini. BDH is said to be inspired by how the human brain works, enabling continuous, energy-efficient learning without needing to store memory as a separate external component.

In practice, this means the model integrates memory, adaptation and reasoning into a single system instead of bolting them on as additional modules. According to the Pathway team, it is this design, not the parameter count or raw computing power, that determines how much a single act of model "reasoning" actually costs.

AI is expensive today, but that cost comes from the architecture, not from any general law of intelligence - Zuzanna Stamirowska, co-founder and CEO of Pathway

Who is behind Pathway

The project is led by Zuzanna Stamirowska, who previously worked on modeling complex systems and predicting traffic in maritime trade. The founding team also includes Adrian Kosowski, a physicist and computer scientist who earned his PhD at age 20, and Jan Chorowski, who previously worked with Geoffrey Hinton's team. The company also counts Łukasz Kaiser, one of the co-creators of the transformer architecture, among its advisors, lending the project extra credibility in the research community.

Pathway is not a project built from scratch. The company was founded in 2019 under the name NavAlgo and has for years sold real-time data processing solutions to clients such as French postal operator La Poste, freight forwarder DB Schenker, and NATO. Building its own language model is, for the company, an extension of the expertise it gained building data infrastructure rather than a leap into an entirely new business.

Funding round and plans

The startup announced on August 11, 2026 that it had raised an additional $20 million as part of an extended seed round, at a valuation reaching $500 million. In total, Pathway has raised about $30 million in this round from investors including Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital and WS Investment Co. The funds are earmarked for GPU infrastructure, including Nvidia GB300 chips, and for training larger multitask models based on the same architecture.

This sets Pathway's strategy apart from most AI labs, which for the past two years have competed primarily on scale, building ever-larger models on ever-larger GPU clusters. Pathway is going against that trend, betting that competitive advantage in the coming years will shift from raw computing power to efficiency, meaning the ratio of answer quality to the cost of generating it.

What it means for the market

The BDH-CQ result does not mean the model matches the best systems in overall capability. Still, a score of 29.5 percent on ARC-AGI-1 with just 150 million parameters is significant, since it shows that a much smaller and cheaper-to-run model can approach the results of far larger systems on specific logical reasoning tasks. For companies deploying AI at scale, where the cost of a single query matters, that kind of gap translates directly into infrastructure bills.

For Poland's AI scene, this is one of the rare examples of a homegrown team pursuing research into a fundamentally new language model architecture, rather than just a wrapper or application built on top of existing OpenAI, Anthropic or Google models. A $500 million valuation on just $30 million raised also shows how much confidence investors are placing in the team, even though the company is still at an early, seed stage.

Pathway's next test will be scaling the BDH architecture to larger multitask models, which the company says it will fund from the current round. Only the results of those larger versions will show whether the cost advantage seen at 150 million parameters holds up once the model has to handle a broader range of tasks than a specialized benchmark like ARC-AGI-1.