Pathway's BDH-CQ Model Cuts AI Reasoning Costs by 11x vs. OpenAI
Researchers at AI company Pathway have introduced a new model called BDH-CQ that utilizes a vector-based cognitive approach to significantly reduce operational costs. Published in August 2026, the study claims the model runs up to 11 times cheaper than OpenAI's entry-level GPT 5.6 Luna, despite achieving comparable performance on nonverbal reasoning benchmarks. The architecture relies on only 150 million parameters, a fraction of the billions used in standard frontier models, suggesting a potential shift toward more efficient 'post-transformer' AI systems.
Key points
- Pathway's BDH-CQ model achieved a 30% score on the ARC-AGI-1 benchmark, solving three out of ten nonverbal reasoning puzzles in two or fewer attempts.
- The model costs approximately 11 times less in token usage than OpenAI's GPT 5.6 Luna, which achieved a slightly higher accuracy score.
- BDH-CQ utilizes only 150 million parameters, contrasting sharply with frontier models like Meta's Llama 3, which employ tens to hundreds of billions of parameters.
- The approach builds on a 2025 precursor model named 'Dragon Hatchling,' which simulated neural connections and strengthening during learning processes.
- Researchers suggest this vector-based cognition could mark the beginning of a 'post-transformer' era, offering a more cost-effective path toward artificial general intelligence.
Background
This development follows recent trends in AI efficiency, including earlier reports on oscillator-based technologies that could be 1,000 times more energy efficient than conventional computing. It also aligns with broader industry efforts to reduce the energy and financial footprint of large language models, as seen in recent expansions of OpenAI's health-focused reasoning capabilities and the development of multimodal pathology models like PRISM2 that prioritize diagnostic accuracy through optimized training methods.
Why it matters
The dramatic reduction in computational cost could democratize access to advanced AI reasoning capabilities, making complex cognitive tasks affordable for smaller organizations and startups. If widely adopted, this vector-based approach could lower the barrier to entry for artificial general intelligence, shifting the industry focus from sheer model size to architectural efficiency and cost-effectiveness.
What to watch
Industry observers will monitor whether other AI developers adopt similar vector-based architectures to compete on cost rather than scale. The next steps likely involve further benchmarking of BDH-CQ against more complex reasoning tasks and potential integration into commercial applications where operational expenses are a primary constraint.
Want the full story? Read the original reporting
Read on Live Science