Model-planning showdown crowns Claude Code

This reads less like a raw model leaderboard and more like a workflow test: the models that asked better questions, stayed on task, and produced usable specs won. The interesting part is that prompting and sampling choices moved Qwen enough to change its standing materially.

• –Claude Opus 4.6 led the pack on the author's ranking, with GLM 5.1 close behind and decent spec depth
• –Qwen 3.6 improved noticeably when preserve-thinking was on and temperature was lowered, which is a reminder that inference settings can matter as much as model choice
• –The local Gemma runs were notably weak, especially the 31B variant that finished after only one question, suggesting poor planning discipline rather than just lower raw capability
• –The setup is methodologically limited, but the "write the spec outside the project tree" constraint is a smart way to reduce self-copying and make planning quality more visible
• –For feature-planning work, this kind of eval favors models that can interview well and structure requirements, not just write fluent prose