It's Not The Tokens Honey It's The Model

In coming to understand AI and its (local!) implementation, we have to not only de-bug but de-myth its marketing nomenclature: mo’ tokens mo’ beddah. But this just isn’t true. Just look at this one qualification:

Qwen3.6 was much faster at generating tokens, but Qwen3.8 reached a usable final answer sooner in the simple comparison because it generated far fewer reasoning tokens.

In a simple three-sentence test, Qwen3.6 generated at roughly 59.5 tokens/sec, versus about 13.4 tokens/sec for Qwen3.8. But 3.6 produced 1,080 completion tokens, of which 994 were reasoning tokens. Qwen3.8 produced only 197 completion tokens, with just 126 reasoning tokens. So despite being much slower per token, Qwen3.8 finished the whole request in about 14.7 seconds, compared with about 18.1 seconds for 3.6.

The same pattern showed itself in the agent-planning test. Qwen3.6 used 1,558 reasoning tokens, while Qwen3.8 used 452. And 3.8’s answer was not merely shorter: it was also more operationally disciplined about failure handling, approval boundaries, and when to stop. Something critically important in AI/agentic workflows. So the practical takeaway was:

3.6 is the faster token engine, but 3.8 was the more reasoning-efficient model.

That efficiency can make it faster in wall-clock time for some tasks despite its lower raw token throughput.