Chinese AI Pricing Cheatsheet

One table to compare Chinese-model prices at a glance

Pricing for Chinese large language models is scattered across provider consoles, English aggregators and forum threads, often in different units and currencies. This cheatsheet pulls the figures into one place and one format: chat models in USD per 1,000,000 tokens, split into input (your prompt) and output (the model's generation), alongside the context window each one offers. The chat table covers the families most teams actually shortlist — DeepSeek, Qwen (通义千问), Doubao (豆包), Kimi, GLM (智谱) and MiniMax — so you can see in seconds where a model sits relative to its peers, whether you are chasing the lowest per-token cost or weighing a larger context window against a higher rate. Because input and output are billed separately, the table keeps them side by side: a model that looks cheap on input can be expensive once a long, verbose answer is generated, and the two columns make that trade-off obvious before you commit.

Image and video generation work differently. Text-to-image models such as Seedream, Qwen-Image, 通义万相 Wan, 混元 HunyuanImage, 文心 ERNIE-Image and 可灵 Kling Image, and text-to-video services such as 可灵 Kling, 即梦 Seedance, 海螺 Hailuo, Vidu and Wan, are billed per image, per credit or by subscription — rates that swing with resolution, duration, frame rate and speed. A single dollar figure would be misleading there, so those tables surface what actually helps you shortlist: the provider, the capability the model is known for, and how you get access (an API, an app, open weights, or a free, paid or credit tier). You narrow the field here, then open the provider's page to confirm the current per-unit price for the exact spec you need.

Estimates, not quotes — always confirm with the provider

The chat prices and image rows were checked against each vendor's own pages; the video prices are still provisional aggregator figures. Each table shows its own date. Provider prices move — with model launches, promotions, exchange rates and tiered or batch discounts — so the figure that was right last quarter may not be right today. That is why the cheatsheet leads with the date and labels unverified entries as "—/待核实" (pending) rather than guessing: a missing number is more honest than a wrong one. When a price matters to a decision, the workflow is simple — shortlist here, then open the provider's own pricing page and confirm the exact rate for your region, currency and usage tier.

Used that way, the cheatsheet is genuinely fast. Scan the chat table to find the price band that fits your volume, glance at the context column to rule out models too small for your documents, then jump to the image or video table to shortlist by capability and access tier. It runs entirely in your browser, reading from the same shared model-data file that powers our other AI cost tools, so the numbers stay consistent across the suite and update in one place. Nothing you do is uploaded, no model is called, and there is no account or limit — it is simply a clean, current snapshot to start a pricing decision from, not the final word on it. One more habit worth keeping: when two models land in the same price band, the cheaper sticker price is rarely the whole story. Check the context window against the length of the documents you actually send, then weigh the output rate against how verbose each model tends to be, because a chattier model at a lower output price can still cost more per finished task than a terse one that charges slightly more. The tables are arranged to make exactly those comparisons quick, so the few seconds you spend here can save a far larger surprise on the monthly invoice later.