Our Kona EBM a 96% vs. 2% Sudoku Benchmark (logicalintelligence.com)

A recent demo by the AI company Logical Intelligence showcased their Sudoku-solving model, Kona, which achieved an impressive 96.2% success rate in solving Sudoku puzzles in real-time, compared to a mere 2% success rate among leading large language models (LLMs) like GPT-5.2 and Claude Opus 4.5. While traditional LLMs struggle with Sudoku due to their autoregressive architecture—which generates solutions token-by-token and cannot effectively revise past decisions—Kona employs a unique reasoning method. It evaluates the entire candidate grid against constraints simultaneously, allowing it to efficiently identify and correct errors without starting over. This breakthrough highlights a significant limitation of current LLMs when it comes to tasks that require spatial reasoning and constraint satisfaction, raising concerns about their applicability in more complex domains such as industrial control systems. The stark contrast between Kona's approach and that of LLMs is not just about solving Sudoku; it indicates foundational architectural differences that may hinder LLMs from tackling similarly structured problems in practical scenarios. Moreover, the cost efficiency of Kona—solving thousands of puzzles for just $4 in compute—further exemplifies the economic advantages of advanced reasoning architectures over traditional LLMs, which incurred costs nearing $11,000 for their unsuccessful attempts.