What this is

This week on Reddit, a developer ran Alibaba's new Qwen3.8-27B on a 22GB RTX 2080TI and surfaced a counterintuitive result: switching reasoning_effort — the dial that tells the AI to "think one more step" before answering — from "xhigh" down to "medium" caused the model's "thinking tokens" (the internal reasoning text it produces before the final answer) on the same task to collapse from 15,000–20,000 to under 2,000. The quality of the final HTML output dropped off a cliff with it — for a small task like Flappy Bird, medium mode barely produced anything playable.

In other words, today's open-source reasoning models (AIs that "think" before answering) have turned "thinking intensity" into a tunable parameter, but users haven't yet internalized its cost implications:

  • For the same task, raising reasoning intensity can swing token consumption by 20x.
  • Token consumption is driven by how much the model "thinks," not by the length of the visible output.
  • If you're billed per token, deployment cost follows your settings — not the user's actual needs.

How the industry sees it

The bullish take: the open-source community is celebrating that Qwen3.8-27B can run a 100K context window (how much text the model can "read" at once) on a consumer-grade 22GB GPU at 40 tk/s (tokens/second). For the first time, local reasoning-model deployment is approaching "actually usable."

The skeptical / risk take:

  • Parameters like reasoning_effort make cost unpredictable. The original poster admits, "I don't know if this is normal behavior or a bug" — meaning even the developer can't properly account for the bill.
  • Under cloud API pay-per-use billing, cranking up reasoning can easily blow out a monthly budget, and no vendor publishes a standard "quality vs. consumption" curve across the different dials.
  • Benchmark leaderboards are typically scored at the model's highest reasoning setting. Enterprises that select models based on these scores will see their budgets and actual consumption diverge sharply.

Impact on regular people

  • For enterprise IT: When selecting a reasoning model, benchmark scores are only the starting point. The reasoning-dial setting is the real cost variable. Budget estimates must factor in "thinking volume."
  • For individual professionals: When using AI for routine coding and writing, lowering reasoning does save time and money — but know that quality drops with it. Don't cut corners with low/medium settings on critical deliverables.
  • For the consumer market: Expect more tiered pricing based on "thinking volume." Regular users will, for the first time, pay for "how much the AI thought," not "how much the AI said."