Claude Opus 4.5 beats human at custom wrap-around crazyhouse chess variant
Defiant_Ranger607 · reddit · 2026-09-30
The poster designed a custom chess variant — the board wraps from the h-file to the a-file, knights move three-then-one, and captured pieces can be dropped back as in crazyhouse — and played it against Claude Opus 4.5. Given only the rules, starting position, and a board diagram, the model produced legal moves each turn and won. It also won at a recreated obscure board game, Nika.
This challenged the poster's prior assumption that chess exposes fundamental LLM limits (tracking a changing board, exact rule-following, planning). They ask the community: which broad problem classes remain unsolvable for LLMs, which limits are architectural rather than improvable, and what would constitute a good test?
More from Models
- OpenAI Codex now natively runs open models like GLM-5.3 Flash and Kimi K3 — HarveenChadha · 2026-09-30
- Cisco's open-weight Antares model localizes vulnerabilities, rivals models 100x larger — aminkarbasi · 2026-09-30
- OpenAI clarifies: chatting with Dots is free, but tasks in Codex count toward limits — karmay007 · 2026-09-30
- ChatGPT hit by elevated error rates for many users, Polymarket reports — Polymarket · 2026-09-30
- Surge AI argues GDP.pdf benchmark training transfers beyond PDFs — echen · 2026-09-30
- GPT-6.1 rolls out to API and all paid plans, pitched as remarkably efficient — TheMoonMidas · 2026-09-30