# Claude Opus 5.5 leads my nonogram benchmark's Hard mode: 8 of 10 random 20x20 puzzles, while 11 of 15 models solve not even one — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/a4a8232e99fde1eb

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-09-30 14:07 ingested 2026-10-01 02:38

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1wu5w1c/claude_opus_55_leads_my_nonogram_benchmarks_hard/)

## retained text



```
I've been benchmarking LLMs on nonograms (picross logic puzzles) since January: one attempt per puzzle, no tools, answers checked against the row and column clues.

On the new Hard mode (ten random 20x20s), Opus 5.5 solves 8 of 10 and is the current #1. 11 of 15 models couldn't solve any of the 20x20s:

- Claude Opus 5.5 (high): 8/10
- GPT-6.1 Sol (max): 7/10
- GPT-6 Astra (xhigh): 5/10
- Claude Fable 5.1 (high): 5/10
- The other 11 models: 0/10

New this week, Claude Sonnet 5.5: 24/30 on Standard mode at max, against 28 for Opus 5.5 at high. Its tokens cost half as much, but it thought so much longer that both runs ended up costing about the same ($5.30 vs $5.24). At low effort it often skipped thinking on the 5x5s and guessed.

Fun detail: Fable did worse at xhigh (1/10) than at high, because 6 of its answers ran into the output limit.

Every answer is public. The puzzle explorer overlays each model's solution on the puzzle: https://www.nonobench.com/puzzles

You can also ask Claude about the results: Nonobench is in the Claude Connectors Directory (https://claude.ai/directory/nonobench).

All results: https://www.nonobench.com
```



document fetched 2026-10-01 00:24 · canonical https://reddit.com/r/ClaudeAI/comments/1wu5w1c/claude_opus_55_leads_my_nonogram_benchmarks_hard
