This week's simple test gave us a wake-up call: Alibaba's Qwen series new model Qwen3.8-27B ran the same task under three "reasoning effort" levels (settings that make the model "think one more step" before generating) — drawing an SVG of "a pelican riding a bicycle." The result: the highest tier xhigh took 717 seconds, 6.4x the lowest tier low (111 seconds), and the visual score only rose from 21.8 to 24.0 (out of 25). A 10% quality bump costs 7x the compute.

What this is

Qwen3.8-27B is a mid-size open-source model from Alibaba's Qwen series, runnable locally on consumer-grade GPUs (such as RTX 5080 laptops with 16GB VRAM). The so-called "three-tier reasoning effort" refers to switches that make the model "think" to varying degrees before answering — low barely thinks and answers directly, while xhigh repeatedly deliberates internally before outputting. The prompt used in this test required the model to return only the standard SVG code for "a pelican riding a bicycle," with no explanation.

Industry view

Supporters say the adjustable-tier mechanism is an engineering advance — complex math and code generation use xhigh, simple Q&A uses low, and users pay per scenario. The dissent splits into two camps: first, the test itself isn't rigorous (only 3 seeds, a single task, third-party scoring) and shouldn't be treated as conclusive; second, even so, the "diminishing marginal returns" issue the test exposes is real. If product managers get used to telling AI to "think at maximum," per-call API costs could 7x while users perceive less than 10% quality improvement — the cost black hole in the middle is the real problem for AI adoption in 2026.

Impact on regular people

For enterprise IT: API billing models need redesign — tier-based differential pricing, or default-to-low with opt-in upgrades, otherwise cloud bills spiral out of control.

For individual workers: when using "deep thinking" modes in ChatGPT, Claude, or ERNIE, judge which tasks are truly worth the wait — rewriting an email probably doesn't need xhigh.

For consumer markets: future AI subscriptions will likely tier out (basic vs. deep-thinking editions), the price gap will widen, and free users may only get the low tier.