Qwen3.6-27B tops Sonnet 4.6 planning

Interesting takeaway: for codebase-aware planning, bigger does not automatically mean better. A smaller dense model can outperform a frontier model if it spends more effort on context inspection and less on confident but generic synthesis. The strongest praise here is for repository fit: Qwen reportedly understood the existing code and integration points better than Sonnet 4.6. The `search_and_read()` suggestion is a useful signal that the model was optimizing the workflow, not just drafting a plan. Sonnet still got credit for access-control and tool-parsing concerns, so the gap looks narrower on general correctness than on detailed system awareness. The comparison is not perfectly controlled because Pi and Claude Code are different harnesses, so some of the outcome may be tooling, not just model weights. If this holds up across more repos, Qwen3.6-27B looks like a strong option for local planning tasks where careful rereads matter more than raw scale.