GPT-6 Astra is the first model to solve all 30 puzzles in my nonogram benchmark

GPT-6 Astra is the first model to solve all 30 puzzles in a nonogram benchmark, which has been evaluating LLMs since January. In the new Hard mode, consisting of ten random 20x20 puzzles, Astra solved 5 out of 10. It successfully completed all five puzzles solvable line by line but failed on those requiring deeper search. Claude Opus 5.5 currently leads this mode with 8 out of 10 puzzles solved, while 11 of 14 models couldn't solve any of the 20x20s. The benchmark is public and open source at nonobench.com.