China’s AI Models Are Rewriting The Cost Of AI

The cost of running many AI workloads is falling, but the bigger change may be in how companies decide which model to use.

DeepSeek, Alibaba's Qwen and Moonshot AI's Kimi have put open-weight models into the market at prices that, in some cases, are a fraction of those charged by leading US model providers. DeepSeek's V4.1 Flash, for example, costs $0.15 per million cache-miss input tokens and $0.60 per million output tokens during off-peak hours. At peak hours, the prices rise to $0.30 and $1.20. OpenAI's GPT-5.6 Sol is priced at $4 per million input tokens and $20 per million output tokens, while Anthropic's Claude Sonnet 5 costs $2 and $10 respectively.

At Alibaba Cloud's international pricing tier for requests with up to 32,000 input tokens, Qwen3-Coder-Next costs $0.30 per million input tokens and $1.50 per million output tokens. Its listed input and output prices at that tier are roughly 13 times lower than GPT-5.6 Sol's. The price rises to $0.50 and $2.50 for inputs between 32,000 and 128,000 tokens, and to $0.80 and $4 for inputs between 128,000 and 256,000 tokens.

The comparison changes when other Chinese models are included. Moonshot's Kimi K3 costs $3 per million cache-miss input tokens and $15 per million output tokens. That is 25% below GPT-5.6 Sol on list prices, but higher than Claude Sonnet 5's $2 and $10.

The pricing data doesn't show that every Chinese model is cheaper than every Western model. It is that developers now have a much wider choice of models at very different price points, with different trade-offs between capability, cost and deployment.

That is changing the economics of AI software.

The Price Of Inference

DeepSeek has been one of the clearest examples of how AI inference costs are being pushed lower.

When DeepSeek-R1 was released in January 2025, the company said it performed on par with OpenAI's o1 and released the model and code under the MIT licence. Its API was priced at $0.55 per million cache-miss input tokens and $2.19 per million output tokens.

The company's latest Flash model shows how the efficiency equation has evolved.

DeepSeek-V4.1-Flash has 552 billion parameters in a mixture-of-experts architecture, but DeepSeek says only 8 billion parameters are active for input and 16 billion for output. The company also says the model's key-value cache requires one-fourth the high-bandwidth memory and one-eighth the SSD storage of the previous generation. It has a 1-million-token context window and supports tool calls.

Those specifications matter because the cost of running an AI model is not determined simply by the total number of parameters. How much compute is activated, how memory is used and how efficiently a model handles context can all affect the cost of serving a request.

Alibaba's Qwen3-Coder-Next is pursuing a similar trade-off for coding. Qwen describes it as an open-weight model designed for coding agents and local development. It uses 3 billion activated parameters out of 80 billion total and has a 256,000-token context window. Qwen says the model delivers results comparable to Claude Sonnet on agentic coding and other coding tasks.

Kimi K3 takes the architecture story in another direction. Moonshot says the model has 2.8 trillion parameters, uses a mixture-of-experts architecture with 16 of 896 experts activated at a time, supports a 1-million-token context window and delivers about a 2.5 times improvement in scaling efficiency over Kimi K2. Moonshot says Kimi K3 has 2.8 trillion parameters, with 104 billion activated during inference.

The models are therefore competing on more than list price. Architecture, context length, inference efficiency, output speed and the amount of computation required to complete a task are increasingly part of the same commercial equation.

That is also why token price alone can be misleading.

Artificial Analysis puts DeepSeek V4.1 Flash's blended token price at about $0.18 per million tokens, compared with $3.08 for GPT-5.6 Sol in its comparison methodology. More importantly, it estimates a cost per benchmark task of about $0.27 for DeepSeek V4.1 Flash versus $1.99 for GPT-5.6 Sol at maximum reasoning effort. The methodology incorporates input, cache, reasoning and output token costs across its benchmark suite.

The difference matters because enterprises ultimately care about the cost of completing a task, not the token price in isolation.

Cheap tokens do not automatically translate into cheap completed tasks. The model's output length, reasoning effort, caching, speed and the workload itself can all change the economics.

That makes the AI price competition less straightforward than a race to the lowest API rate.

From One Model To A Model Stack

The commercial effect is visible in how developers and enterprises are accessing these models.

Amazon Web Services added DeepSeek V3.2, Kimi K2.5 and Qwen3-Coder-Next to Amazon Bedrock in February, describing them as open-weight models that offered frontier-class performance at significantly lower inference costs. Kimi K3 became generally available on Bedrock on September 18. AWS said it brings a 1-million-token context window to the platform and runs within the same security and governance boundary as proprietary models.

That means an enterprise does not necessarily have to choose between buying a proprietary model through an external API and building an AI system from scratch. It can access an open-weight model through a cloud provider, use a compatible hosted service or, subject to licensing and infrastructure requirements, deploy the model on its own systems.

The models are also appearing inside the tools developers already use.

DeepSeek's API documentation lists integrations with Claude Code, GitHub Copilot and OpenCode. Qwen3-Coder-Next is designed to work with Qwen Code, Cline and Claude Code. GitHub added Kimi K3 to GitHub Copilot in August and said the model is billed at provider list pricing under usage-based billing.

The competition is therefore no longer just about which chatbot a consumer opens.

A coding assistant can use one model. A document-processing workflow can use another. A company may use a lower-cost model for summarisation and extraction while reserving a more expensive frontier model for complex reasoning.

OpenRouter, a platform that routes requests across multiple model providers, offers a view of how developers are making those choices. In the week beginning September 14, DeepSeek accounted for 25.4% of text requests on OpenRouter, while Qwen accounted for 6.7%. OpenRouter's figures measure request share on its own platform and do not represent the global AI market.

DeepSeek was the largest model author by share of requests on OpenRouter that week, ahead of Google and OpenAI. The data shows that Chinese model providers are gaining meaningful developer traffic on a platform built around model choice. It does not establish their share of global AI usage.

There is evidence of the same model-choice approach inside enterprises.

Siemens says its developer infrastructure supports a range of open-weight models, including Qwen and DeepSeek, alongside Llama, Mistral and others. The company said its internal infrastructure processed millions of requests, hundreds of millions of generated tokens and billions of prompt tokens in a single month. The figures cover the overall model infrastructure rather than Qwen and DeepSeek alone, but Siemens said those models were among those available to its developers.

Notably, the data don't point that enterprises are choosing Chinese models instead of proprietary ones. It is that the model itself is becoming a more modular part of the software stack.

Open Weights Change The Equation

Open-weight models add another layer to that modularity.

DeepSeek V4.1 Flash is released under the MIT licence, while Qwen's open-weight models, including Qwen3-Coder-Next, are distributed under Apache 2.0. Kimi K3 uses a separate Kimi K3 licence. Its terms allow use, modification, distribution and deployment, but impose additional conditions on qualifying Model-as-a-Service businesses above $20 million in aggregate revenue over a 12-month period and on some commercial products exceeding 100 million monthly active users or $20 million in monthly revenue.

That distinction matters because downloadable model weights can give companies more deployment options. But open weights do not make inference free.

Running a large model still requires accelerators, memory, electricity, infrastructure and engineering. Moonshot recommends deploying Kimi K3 on supernode configurations with 64 or more accelerators.

The economics also vary sharply by workload.

Kimi K3 is a useful counterexample to the idea that every Chinese model is simply cheaper. Its list price of $3 per million cache-miss input tokens and $15 per million output tokens is below GPT-5.6 Sol's $4 and $20, but above Claude Sonnet 5's $2 and $10. For the max-effort configurations in Artificial Analysis's benchmark comparison, Kimi K3 comes in at about $2 per task versus $1.99 for GPT-5.6 Sol. The figures are based on the evaluator's task mix and do not represent the cost of every real-world workload.

Cheap tokens, in other words, do not automatically produce a cheaper AI system.

The same is visible in model-download activity. Hugging Face recorded more than 570,000 downloads of DeepSeek V4.1 Flash and more than 520,000 downloads of Qwen3-Coder-Next in the previous month. At the time of publication, Hugging Face showed more than 1.7 million downloads of Kimi K3. These figures measure downloads from model repositories, not confirmed production deployments.

For developers, this creates a different decision from the one that dominated the early generative-AI market, when the focus was largely on identifying the strongest general-purpose model.

Now the questions are more granular.

Which model can complete a task at the lowest cost? How much reasoning does it require? Where does the model run? What does the licence allow? How much context does the application need? And does the model's performance justify the cost of running it?

Those questions are pushing AI deployment towards a multi-model architecture.

The result is a market where the advantage does not necessarily go to the model with the highest benchmark score or the lowest token price. It can go to the model that gets a specific job done at a cost a business can support.

Chinese open-weight models have become an important force in that shift. They have contributed to downward pressure on API prices, widened access to capable open-weight models and expanded the range of price-performance combinations available to developers and enterprises.

The competitive pressure is no longer only about who has the best model. It is increasingly about who can provide the right combination of capability, inference cost and deployment flexibility for a given task.