AnalysisAI ModelsSeptember 9, 2026

Qwen3.8-27B defaults to xhigh reasoning effort, costing 8x tokens

Qwen3.8-27B's chat template sets reasoning_effort to xhigh by default, the most expensive of its three settings. On one Mac, xhigh used 14 minutes per question versus 77 seconds for low/medium, moving median coding score only from 78.2 to 78.8.

1 source

More stories today

Open the live feed