DeepSeek and other Chinese AI: which models to choose and how much they cost

DeepSeek

The Hangzhou-based company offers two main models — DeepSeek-V4-Flash and DeepSeek-V4-Pro. Both support a context window of one million tokens, a maximum response size of up to 384,000 tokens, and a switchable reasoning mode.

V4-Flash costs $0.22 per one million input tokens and $0.66 per one million output tokens. When hitting the internal cache, the price for input tokens drops to $0.007 — this is the cheapest option among all models in the review. V4-Pro costs $0.66 per one million input tokens and $1.98 per one million output tokens.

The model supports tool calling, structured JSON output, and works through interfaces compatible with OpenAI and Anthropic. In early August 2026, DeepSeek warned developers about an impending significant price increase, without disclosing exact figures or dates.

Who it is suitable for: developers who prioritize the lowest possible input token price for large volumes of requests — text classification, data extraction, batch processing.

Qwen from Alibaba Cloud

The flagship Qwen line is available via the Alibaba Cloud Model Studio platform, formerly known as DashScope. The Qwen3.7-Max model, with a context window of one million tokens, is priced at $2.5 per one million input tokens and $7.5 per one million output tokens under the current promotion — compared to the base price of $5 and $15, respectively.

The budget-friendly Qwen-Flash line starts from $0.05 per one million input tokens and $0.4 per one million output tokens for requests up to 256,000 tokens — prices increase for longer requests. The specialized coding model Qwen3-Coder-Plus costs from $1 per one million input tokens up to $6 for requests longer than 256,000 tokens.

An important feature of Alibaba Cloud is that prices depend on the deployment region. Rates for the Singapore data center, which most developers outside China use, are noticeably higher than those within China: for example, Qwen-Plus in Beijing costs approximately three times less than in Singapore.

Some models in the family — including Qwen3, Qwen3.5, and Qwen3-Coder — are released with open weights under the Apache 2.0 license, allowing them to be deployed on your own servers without paying per token.

Who it is suitable for: teams needing flexibility between cloud API and local deployment, as well as projects with multimodal tasks — Qwen has separate models for image, audio, and video recognition.

Kimi from Moonshot AI

Beijing-based Moonshot AI develops the Kimi line. The current flagship, Kimi K3, was released on July 16, 2026 — it is a 2.8 trillion parameter model with a context window of 1,048,576 tokens, constantly operating in reasoning mode with adjustable depth via the reasoning_effort parameter.

According to the platform's official documentation, K3 costs $3 per one million input tokens on a cache miss, $15 per one million output tokens, and only $0.30 per one million input tokens on a cache hit. Moonshot claims that for typical development tasks, the cache hit rate exceeds 90%, meaning the real cost of input tokens in practice is much closer to $0.30 than to $3.

More affordable models in the line — Kimi K2.6 and the specialized coding model Kimi K2.7 Code — cost $0.95 per one million input tokens and $4 per one million output tokens with a context window of 262,144 tokens. The weights for K3 were published openly on July 26, 2026, enabling self-deployment if you have the appropriate infrastructure.

Who it is suitable for: teams working with large codebases and agent scenarios where the model repeatedly accesses the same context — this is where caching provides the greatest savings.

GLM from Z.ai

Beijing-based Zhipu AI, operating in the international market under the brand Z.ai, offers the GLM line. The current flagship model GLM-5.2 costs $1.4 per one million input tokens and $4.4 per one million output tokens, while cached input tokens cost only $0.26 per one million. The model supports a context window of up to one million tokens.

The earlier GLM-4.6 costs $0.6 per one million input tokens and $2.2 per one million output tokens, while the lightweight GLM-4.5-Air costs $0.2 and $1.1, respectively. Separately, Z.ai offers two completely free models — GLM-4.7-Flash and GLM-4.5-Flash — available via API without token fees, though with limitations on speed and reasoning depth.

Like Qwen, the flagship GLM models are released with open weights under the MIT license, allowing self-deployment even for top-tier versions — a rare case among large model developers.

Who it is suitable for: projects with limited budgets, where some tasks can be handled by free models, while more complex ones are directed to the paid flagship.

What to keep in mind

Prices on all listed platforms change frequently. Before starting work with any of the models, it is worth checking the official pricing page rather than relying on figures from an article written three months ago.