Skip to content
Hot eventDeveloping

Logical Intelligence benchmark pits Kona against LLMs on Sudoku

1 reports1 reporting sourcesUpdated 50 min ago

Start here

AI digest

Logical Intelligence says its energy-based model Kona solved 96.2% of hard Sudoku puzzles in an average of 313 milliseconds, while GPT-5.2, Claude Opus 4.5, Sonnet 4.5, Gemini 3 Pro and DeepSeek V3.2 together solved only 2%, with GPT-5.2 best at 6.9%. The company attributes the gap to autoregressive token generation, which it says cannot revise earlier decisions; Kona instead scores a complete candidate grid by energy and uses gradient information to repair violated constraints. The comparison was run with code execution and fine-tuning disabled for all models, and Logical Intelligence says Kona was not trained on solved puzzles. It reports roughly $4 in GPU time for Kona across nearly 13,000 solved puzzles versus about $11,000 in LLM API calls. These figures and the solve rates are the company's own reported results.

Generated from reports · updated 26 min ago

Report timeline

Follow the reports to see the event from different sides.

November 18
  1. Hacker News · AI
    Logical Intelligence pits energy-based model Kona against GPT-5.2 and Claude on sudoku

    Logical Intelligence reports that its energy-based model Kona solved 96.2% of hard sudoku puzzles in an average of 313 milliseconds, while GPT-5.2, Claude Opus 4.5, Sonnet 4.5, Gemini 3 Pro and DeepSeek V3.2 together solved only 2%, with GPT-5.2 best at 6.9%. The company says LLMs fail because autoregressive token generation cannot revise earlier decisions, whereas Kona scores a complete candidate grid by energy and uses gradient information to repair violated constraints. It reports roughly $4 in GPU time for Kona across nearly 13,000 solved puzzles versus about $11,000 in LLM API calls, with code execution and fine-tuning disabled for all models.

Heat for this event

Not enough continuous observations yet to draw a trend.