Chinese AI API Pricing

Short answer. Chinese API list prices have converged on a shape: input around $0.15 for a flash tier, output at 3–5x input, and a cache hit at 2–20% of input. What differs — and what actually decides your bill — is the structure around the price: time-of-day tiers, context cliffs, cache residency charges, batch availability and the currency you are invoiced in.

The $0.15 floor is real, and it is not the whole story

Six providers, and almost every one has a model at or near $0.15 per 1M input tokens. That number has become the market's anchor, and it is genuinely available — Qwen3.8-Flash, GLM-5.3-Flash and DeepSeek-V4.1-Flash all sit there. But the floor is a floor on input, and input is rarely the dominant cost.

Output is where bills are made. At a 4:1 ratio — the median here — a workload that generates 500 tokens for every 100 it reads spends 95% of its money on output. Comparing two models at the same input price but different output ratios is comparing the smaller half of the bill.

Cache depth varies by a factor of ten

The cache discount is the second lever, and it is the one buyers under-use. A cache hit costs 2% of input at DeepSeek, 10–12% at Moonshot and Alibaba, and 20% at ByteDance, MiniMax and Zhipu. On an agent loop that re-reads a 40K-token system prompt twenty times, that difference is larger than the gap between the headline prices.

The catch is that caching rewards stable prefixes. Prompts that reshuffle context on every call — the common RAG pattern — get little benefit from any provider, and should be optimised before the provider is chosen.

Structure beats list price

• Time-of-day pricing (DeepSeek): the cheap rate is available to queues, not to chat. A European product's working day overlaps the peak window.
• Context cliffs (MiniMax): above 512K input tokens the whole call doubles, so a 1M-context model is not flat-priced across its window.
• Tier spreads (Zhipu 9x, Moonshot 20x): the same family sold at wildly different prices, which makes routing — not provider choice — the biggest cost lever.
• Cache residency (ByteDance): the only provider charging for cache storage per hour, which punishes caches that are created and forgotten.
• Currency (ByteDance, CNY only): the dollar cost is an FX decision, and the invoice may not fit your finance system.

Read the table below with three questions

First, what does a completed task cost at your output length — not at your input length. Second, what fraction of your prompt is a stable prefix that can be cached. Third, which tier boundary your workload crosses: the 512K context cliff, the batch availability gap, or the peak window.

Answer those three and the per-token price becomes almost irrelevant. Ignore them and the cheapest-looking model on this page is frequently the most expensive one to operate.