Tiny AI model 11 times cheaper than GPT-5.6 Luna and almost matches it in logic test

Pathway Lab, which focuses on architectures different from transformers, released results for its reasoning model BDH-CQ. Despite having only 150 million parameters, the model achieved 29.5% on the ARC-AGI-1 benchmark, while costing just $0.0007 per task. That makes it 11 times cheaper than OpenAI's small model GPT-5.6 Luna, which scored 34.2% and cost $0.008 per task even after an 80% price cut. For comparison, Anthropic Claude Opus 5 and Google Gemini 3.1 Pro scored 97–98% on the same test but cost $0.5–$0.6 per task, almost a thousand times more. Alibaba Qwen3 235B cost more than three times as much as BDH-CQ but performed worse. The lab says the difference comes from reasoning mechanisms: most current AI models generate intermediate text that consumes tokens, while BDH-CQ solves problems silently in memory. Initial experiments showed that BDH-CQ follows standard scaling laws comparable to transformer models with 1 to 600 billion parameters. Pathway plans to tackle harder tasks, including mathematical reasoning, the ARC-AGI-2 test, and eventually a full evaluation of ARC-AGI-3, hoping the efficiency advantage holds on larger and more complex problems.