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.
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.
| Model | Tier | Input (non-cached) | Input (cached) | Output |
|---|---|---|---|---|
| GLM-4.6 (355B A32) | Input ≤32k · Output ≤200 | 1 | 0.2 | 4 |
| Input ≤32k · Output >200 | 1.5 | 0.3 | 7 | |
| Input 32k–200k | 2 | 0.4 | 8 | |
| Qwen3-Max | Input 0–32k | 1.6 | 0.32 | 6.4 |
| Input 32k–128k | 3.2 | 0.64 | 12.8 | |
| Input 128k–252k | 4.8 | 0.96 | 19.2 | |
| Qwen-Plus (235B A30) | 0–128k · non-thinking | 0.4 | 0.08 | 1 |
| 0–128k · thinking mode | 0.4 | 0.08 | 4 | |
| Qwen-Flash | Input 0–128k | 0.075 | 0.015 | 0.75 |
| Input 128k–256k | 0.3 | 0.06 | 3 | |
| GLM-4.5V / 4.5-Air | Input 0–32k | 1 | 0 | 3 |
| Input 32k–64k | 2 | 0 | 6 |
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.
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.”
Qwen-Plus, GLM-4.5-Air
Balanced value. The daily workhorse for chat, RAG Q&A, and summarization—most requests should land here.
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.
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.
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.