Post-Transformer AI Model Cuts Reasoning Costs by 11x

A new AI model architecture could fundamentally reshape the economics of artificial intelligence by slashing the computational costs of complex reasoning tasks.

Researchers at AI company Pathway have developed BDH-CQ, a model that scored nearly 30% on the ARC-AGI-1 benchmark — a standard test for measuring progress toward artificial general intelligence using nonverbal reasoning puzzles. While that score trails some leading models, the breakthrough lies in efficiency: BDH-CQ achieved its results at roughly one-eleventh the token cost of OpenAI's GPT 5.6 Luna (Low) model, which scored only slightly higher.

The findings were detailed in a research paper published August 10 on the preprint server arXiv.

How post-transformer architecture works

Most contemporary AI systems, including ChatGPT and Claude, rely on transformer architectures that process prompts by predicting the next word in a sequence. These models essentially verbalize their reasoning step-by-step, consuming tokens — the metering units AI companies use to measure computational costs — with each word generated.

BDH-CQ takes a different approach. Instead of storing conversation histories as text strings, it uses numerical arrays to represent underlying rules and contextual patterns. These vectors point to relationships between information chunks stored in GPU memory during training, allowing the model to process abstract reasoning without expanding its memory footprint.

The model employs what Pathway calls a "latent reasoning engine" — an internal workspace that runs iterative loops to refine answers. Each loop takes the previous output, assesses potential improvements based on training data, and feeds the result into the next iteration. Crucially, running 200 loops consumes proportionally the same memory and power as running 20 loops, whereas transformer models burn exponentially more resources as they generate longer reasoning chains.

The parameter efficiency advantage

BDH-CQ was trained on just 150 million parameters, a fraction of the tens or hundreds of billions used by frontier models like Meta's Llama 3.1 405B. Fewer parameters mean faster training and cheaper operation — advantages that could prove decisive as AI deployments scale.

The researchers suggest their model's cognitive capabilities could expand significantly when scaled to larger parameter sizes, potentially maintaining the cost efficiency while improving performance.

Transformer models face inherent scaling challenges. Their computational complexity increases quadratically with prompt length — doubling input size quadruples processing requirements. They also create memory bottlenecks by re-reading every previous word with each new word generated, eventually overwhelming GPU capacity.

Why it matters

As enterprises deploy AI across operations, computational costs have emerged as a critical constraint. A model that delivers competitive reasoning at a fraction of the expense could accelerate adoption in cost-sensitive applications like cybersecurity incident response and industrial operations, where Pathway plans to focus development. If the architecture proves viable at scale, it could challenge the dominance of transformer-based systems that have defined the current AI generation.

Validation and next steps

Prominent AI researchers have independently verified BDH-CQ's benchmark results, including NYU's Richard Zhong and Łukasz Kaiser, co-author of the foundational 2017 "Attention Is All You Need" paper that introduced transformers to large language models.

"Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning," Kaiser said in a statement.

Pathway plans to scale the architecture to 600 billion parameters and develop a full large language model based on the technology. The company will test the approach against more challenging benchmarks including ARC-AGI-2 and ARC-AGI-3.

The details were first reported by Live Science.