A startup called Pathway has unveiled a 150-million parameter reasoning model that challenges the economics of today's frontier artificial intelligence systems. The model, called BDH-CQ, achieved 29.5% accuracy on the ARC-AGI-1 benchmark (a test of reasoning ability) while costing just $0.0007 per task to run. By comparison, OpenAI's GPT-5.6 Luna, even after an 80% price cut in July, costs 11 times more for only slightly better performance at 34.5% accuracy.

This breakthrough matters because it suggests a fundamentally different architectural approach to building AI models could deliver comparable results at a fraction of the cost and computing power. Pathway's innovation challenges the prevailing assumption that bigger models trained on more data are the only path forward in artificial intelligence.

What Makes Pathway's Architecture Different?

Pathway's post-transformer architecture represents a departure from the transformer design that powers ChatGPT, Claude, and virtually every modern large language model (LLM). An LLM is a type of artificial intelligence trained on vast amounts of text to predict and generate human language. The transformer, introduced in a 2017 Google research paper, became the foundation for the generative AI revolution.

The company's approach, detailed in a paper published in October 2025 titled "The Dragon Hatchling: The Missing Link between the Transformer and Models of the Brain," offers a theoretical framework for building AI systems that require less data and less computing power while maintaining reasoning capabilities. BDH stands for "beautiful dragon hatchling," a reference to Terry Pratchett's science fiction and fantasy work.

"We need to be able to squeeze more intelligence per dollar, and for this you need to change the paradigm. This is a very deep innovation, and we wouldn't have done it if it wasn't going to be, and if it didn't have a chance to capture the market," said Zuzanna Stamirowska, CEO and co-founder of Pathway.

Zuzanna Stamirowska, CEO and co-founder of Pathway

The Pathway team believes their breakthrough represents "a PageRank moment for intelligence," referring to the turning point when Google founders Larry Page and Sergey Brin realized they could make web search dramatically better by ranking pages based on link structure rather than just keywords.

Who Is Behind This Research?

Pathway assembled a team with deep roots in AI research and industry leadership. The company's advisors and team members include several figures with significant track records in artificial intelligence development:

• Łukasz Kaiser: Co-author of the original transformer paper that launched the generative AI revolution, now serving as an advisor and independently verified the ARC-AGI-1 benchmark results
• Adam Kurzrok: Former group product manager of Gemini at Google DeepMind, now chief product officer at Pathway
• Jonathan Frankle: Chief AI scientist at Databricks, serving as an advisor on scaling and deployment, and also an investor in the company
• Martín Farach-Colton: Chair of computer science and engineering at New York University and one of Google's early employees who led important technological breakthroughs

This roster signals serious technical credibility. Łukasz Kaiser's involvement is particularly notable, given that he co-authored the transformer paper that became the foundation for modern AI systems. His independent verification of the benchmark results adds weight to Pathway's claims.

How Does This Compare to Today's Leading Models?

The cost comparison is striking. On the ARC-AGI-1 benchmark, Pathway's BDH-CQ model achieved 29.5% accuracy at $0.0007 per task. OpenAI's GPT-5.6 Luna (Low), the company's most cost-effective model, scored 34.5% on the same test but cost $0.0077 per task, even after OpenAI slashed prices by 80% on July 30.

The performance gap is modest, but the cost difference is enormous. Part of the reason Pathway's model is so much cheaper is that it is smaller (150 million parameters versus much larger frontier models), doesn't require chain-of-thought reasoning (a technique where models work through problems step-by-step), and needs less training data due to improved memory architecture.

Pathway's team believes their architecture can scale to 600-billion parameter models that would compete with the world's leading frontier models while maintaining efficiency advantages. However, the company acknowledges it still has significant work ahead before its models can match the capabilities of today's most advanced systems.

Why Does This Matter for AI's Future?

The AI industry is facing mounting pressure on multiple fronts. Current scaling laws that power today's top models consume enormous amounts of computing power, energy, and data, and many researchers believe these approaches are hitting fundamental limits. The industry desperately needs efficiency breakthroughs to make AI more sustainable and economically viable at scale.

Pathway is one of over 40 startups, collectively called "neolabs," that have raised $40 billion in funding specifically to attack these problems and develop post-transformer architectures. Whether or not Pathway becomes the eventual leader in this space, these startups represent a significant shift in how the AI industry is approaching model development.

How to Evaluate Emerging AI Architecture Breakthroughs

When assessing claims about new AI architectures, several factors matter:

• Benchmark Verification: Look for independent verification of benchmark results by respected researchers. Pathway's use of Łukasz Kaiser to independently verify results adds credibility to their claims
• Team Credentials: Examine whether the team includes researchers and engineers with proven track records in AI development. Pathway's advisors include co-authors of the transformer paper and former Google leaders
• Scalability Path: Assess whether the approach has a clear path to scaling up to frontier-class model sizes. Pathway's roadmap to 600-billion parameter models suggests they have thought through this challenge
• Practical Cost Metrics: Compare not just accuracy but inference cost, training requirements, and data efficiency. Pathway's 11-fold cost advantage over OpenAI's Luna model is a concrete, measurable claim

Pathway's breakthrough is still early stage. The company has demonstrated promising results on a specific benchmark with a relatively small model, but scaling these results to frontier-class performance remains unproven. However, the team's pedigree, the independent verification of results, and the magnitude of the cost advantage suggest this is a development worth monitoring closely.

The broader context matters too. This announcement comes as Google's AI leadership undergoes significant restructuring. Demis Hassabis, the longtime CEO of Google DeepMind, is stepping back to focus on broader AI strategy and his drug discovery company, Isomorphic Labs. Koray Kavukcuoglu, previously Google DeepMind's chief technology officer, is taking over as senior vice president and reporting directly to Google CEO Sundar Pichai.

The leadership change reflects growing pressure on Google to compete more aggressively with OpenAI and Anthropic in frontier AI development. Meanwhile, startups like Pathway are pursuing a different strategy: building more efficient architectures that could reshape AI economics entirely. The next 12 to 18 months will reveal whether these post-transformer approaches can deliver on their promise.