Token Cost Engineering · 3 / 13

Price Sheet at a Glance and the Three Tiers

Spread the prices of the models you use onto one table and three clear tiers jump out. But knowing which is expensive and which is cheap is nowhere near enough—the real cost assassins hide in the price-jump boundaries.

T0 FlagshipT1 WorkhorseT2 High-volumecached price
Price sheet panorama

Below are the model prices the author's team mainly used at the time (already converted to post-discount rates, unit: yuan / million Tokens). The three columns are non-cached input, cached input, and output—note that cached price is usually only a fifth of standard price or even lower; that huge gap foreshadows Lesson 11 on KV Cache.

ModelTierInput (non-cached)Input (cached)Output
GLM-4.6
(355B A32)
Input ≤32k · Output ≤20010.24
Input ≤32k · Output >2001.50.37
Input 32k–200k20.48
Qwen3-MaxInput 0–32k1.60.326.4
Input 32k–128k3.20.6412.8
Input 128k–252k4.80.9619.2
Qwen-Plus
(235B A30)
0–128k · non-thinking0.40.081
0–128k · thinking mode0.40.084
Qwen-FlashInput 0–128k0.0750.0150.75
Input 128k–256k0.30.063
GLM-4.5V / 4.5-AirInput 0–32k103
Input 32k–64k206

Red numbers are prices “after a tier jump.” Inside the same model, prices can differ by 2–3×—those boundaries are the stars of Lessons 4, 5, and 6.

The Three Tiers
T0 · Flagship

Qwen3-Max, GLM-4.6 (long-output tier)

Expensive output, capability ceiling. Leave for complex reasoning, code gen, multi-model arbitration—tasks where “getting it wrong costs more.”

T1 · Workhorse

Qwen-Plus, GLM-4.5-Air

Balanced value. The daily workhorse for chat, RAG Q&A, and summarization—most requests should land here.

T2 · High-volume

Qwen-Flash

Nearly free (cached input 0.015 yuan/M). Fine for data cleaning, intent classification, high-frequency monitoring—and the best sidekick for bigger models.

Tech folks stare at model params; operators stare at pricing ladders. In AI engineering, you need both.
Interactive drill · Which tier for this task

Five real business scenarios—pick the tier you think fits best. Principle: never pay for expensive when cheap works, but don't skimp when being wrong is costly.

Key Takeaways

Tier first, then select: T0 for hard problems, T1 for daily work, T2 for volume. Most cost accidents are “using T0 to do T2's job.”

Cached price is 1/5 of standard—or lower. Whether you get that discount depends on your architecture (see Lesson 11).

Inside one model, prices can differ 2–3×. Tier-jump boundaries (output 200, input 32k) deserve more attention than model selection itself.

Source: Adapted from the author's internal team share “AI Token Cost Engineering Strategies.” Table figures are the author's post-discount contract rates at the time; list prices and discounts change constantly—before selecting, verify live quotes on DeepSeek Pricing, Alibaba Cloud Model Studio (Bailian), and Zhipu's open platform.