Nonobench v1.2: 43 LLMs on nonogram puzzles. Open-weight DeepSeek V4 Pro ties for 4th, and no open model solves the new 20×20 Hard mode

Follow-up to my January post: . That thread shaped v1.2: Reasoning effort is explicit per run Every prompt and output is public. All current top ranking private and open weight models have been added Someone spotted Grok miscounting a 400-character answer. Turns out that trips up most models, so Hard mode answers row by row rather than a single text string. Results: - GPT-6 Astra: 30/30, the first perfect run on 15x15 puzzles - Best open weights: DeepSeek V4 Pro 83% (tied 4th), DeepSeek V4.1 Flash 77% for $0.84 total - Hard mode (10 random 20×20s, one solution each): Opus 5.5 8/10. Every open-weight model: 0/10 Still OpenRouter-only, so no way to run locally yet. PRs welcome. nonobench.com (raw data, API, and code on GitHub)