Controlled comparison
Thinking vs no-thinking: what reasoning costs in a chat
Should you leave the reasoning block on for everyday assistant tasks on a local Qwen3.6-35B-A3B?
What was held fixed
Same unit, backend, artifact and corpora; the only change is the reasoning toggle (chat-template switch, 4096-token budget when on).
| Metric | Reasoning off | Reasoning on (4096) |
|---|---|---|
| Time to first ANSWER token, median (chat) | 210msmedian over valid requests · lab_unit_replicated · evidence | 20191msmedian over valid requests · lab_unit_replicated · evidence |
| End-to-end time to complete JSON answer, median | 844msmedian over valid requests · lab_repeated · evidence | 14677msmedian over valid requests · lab_repeated · evidence |
| Strict-JSON task success | 100.0% of requestsshare of all issued requests · lab_repeated · evidence | 100.0% of requestsshare of all issued requests · lab_repeated · evidence |
How to read this
For these everyday tasks reasoning multiplied the wait by two orders of magnitude and improved nothing the gates can measure: task success is identical with and without it. The user-visible difference is an instant assistant versus a twenty-second pause before the first answer token.
Limits
Everyday-task and strict-JSON corpora only — reasoning may pay for itself on genuinely hard problems (math, multi-step planning) these workloads do not contain. The cost side is general; the benefit side is scoped.
Answering this question somewhere? Paste the table
Markdown, renders on GitHub / Reddit / forums. Attribution line included — the data is CC BY 4.0, take it.