Why are Chinese AI models cheaper?

Why are Chinese AI models cheaper?

Chinese AI models are cheaper because they are built to be cheap to run. Each answer uses a small slice of a very large model, open weights let dozens of companies compete to host them, and the labs price for growth. The cheapest Chinese models cost a thirteenth as much per token as US flagship models, or less.

The gap has changed in 2026. OpenAI and Anthropic now sell budget models at the same per-token price as the cheapest Chinese ones, so the price advantage holds against US flagships and has mostly closed against the US budget tier. The table below shows both.

I'm a very heavy AI user, and my own bill shows what these prices mean in practice. From 16 August to 15 September 2026, two Chinese everyday chat models cost me $18.90 in total: $10.00 on DeepSeek V4 Flash and $8.90 on Z.ai's GLM 5.3 Flash. I use them through OpenRouter, a pay-per-use service, in a desktop chat app that runs on my own API key (a setup for technical users). My view is that cheap models like these sit slightly off the frontier and are great for everyday use.

How much cheaper are Chinese AI models?

List prices per million tokens come from each lab's own pricing page, checked 11 October 2026. "Score" is the Artificial Analysis Intelligence Index, a composite of hard reasoning, knowledge, coding and agent tests, at each model's highest effort setting. "Cost per task" is Artificial Analysis's measured average cost to run a model through those tests. Both were checked the same day.

Sources: DeepSeek API docs, Z.ai, Xiaomi's own endpoint on OpenRouter, Alibaba Cloud Model Studio (international), Moonshot's Kimi platform, OpenAI API pricing, claude.com/pricing and Google's Gemini API pricing, all 11 October 2026. Qwen3.8 Flash is Alibaba's API model; Artificial Analysis scores only its open sibling, Qwen3.8-Flash-Next (40). Claude Haiku 5.5's price applies to prompts up to 100,000 tokens. Gemini 3.8 Flash is listed at $1.50 and $7.50 from 1 January 2027.

Against US flagships, the Chinese efficient models are far cheaper per token. DeepSeek's off-peak input price is a thirteenth of GPT-6.1 Sol's and a sixty-seventh of Claude Fable 5.1's. GLM-5.3-Flash's output price is a twentieth of Sol's and a hundredth of Fable's.

Against US budget models, the gap is gone. GPT-6 Luna and Claude Haiku 5.5 list slightly below DeepSeek's off-peak price, and Haiku 5.5 scores higher than DeepSeek V4.1 Flash.

The Chinese flagships cost about what US mid-tier models do. Kimi K3 lists above GPT-6.1 Sol on both input and output.

How Chinese labs make each answer cheap to serve

Most of what an AI company pays per answer is the computing needed to produce each token. Chinese labs build very large "mixture-of-experts" models, which split the model into many specialist blocks and switch on only a few of them for each token. The model cards on Hugging Face give the numbers:

• DeepSeek V4.1 Flash: 552 billion parameters, with 8 billion active while reading your input and 16 billion while writing the answer.
• GLM-5.3-Flash (Z.ai): 320 billion parameters, 18 billion active.
• MiMo-V2.6-Pro (Xiaomi): 1.02 trillion parameters, 42 billion active.
• Kimi K3 (Moonshot AI): 2.8 trillion parameters, 104 billion active.

So each token DeepSeek writes takes roughly the computing of a 16-billion-parameter model, while the model draws on a far larger pool of parameters. DeepSeek also cuts the memory each conversation takes up: its model card says V4.1 Flash keeps about an eighth of the cached conversation state that V4 Flash did. Input that repeats earlier text and hits the cache is billed off-peak at a fiftieth of the normal input price.

US labs use the same design where they publish the details. OpenAI's open-weight gpt-oss-120b has 117 billion parameters, with 5.1 billion active per token. The closed US flagships don't publish their sizes, so their serving costs can't be compared directly.

Open weights mean dozens of hosts compete on price

Most leading Chinese models are released as open weights under permissive licenses, so anyone can download and run them. Once the weights are public, any cloud company can sell access. On 11 October 2026, OpenRouter listed 29 hosts serving DeepSeek V4.1 Flash, from DeepSeek itself to US clouds such as CoreWeave, Together AI and Fireworks. Most charged for input somewhere between just below DeepSeek's own off-peak rate and its peak rate. A closed model's price is set by the one company that owns it.

I also think China pushes its labs toward open weights. The published explanations, from soft power to setting global standards, are in why China is open-sourcing its AI models.

Pricing for market share, and the subsidy question

Prices also reflect strategy. DeepSeek charges half price outside weekday peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC). Its prices also move quickly. When it introduced peak and off-peak pricing on 16 August 2026, its Flash model's output price more than doubled off-peak and rose almost fivefold at peak. When V4.1 Flash launched on 10 September, it cut the input price by about a third, to the off-peak rate in the table above (DeepSeek's API change log and archived pricing pages).

My view is that Chinese state support helps keep these labs going and their prices low. Public reporting doesn't trace subsidy money to API prices, so read that as my opinion. What is on record: China's government backs its AI sector through state investment funds and its "AI+" initiative, issued by the State Council in August 2025. DeepSeek relied on the Chinese hedge fund High-Flyer for money until June 2026, when it raised more than $7.4 billion in its first outside round, with China's state-backed National Artificial Intelligence Industry Investment Fund investing directly (The Wall Street Journal, June 2026).

Analysts also point to cheap power at data-center hubs in western China, built under the government's "East Data, West Computing" program. The analyses that cite it give no figure for how much of the price gap it explains.

Did US chip export controls make Chinese AI cheaper?

They pushed Chinese labs toward efficiency. The US Commerce Department's Bureau of Industry and Security restricted exports of top AI chips to China on 7 October 2022, then tightened the rules on 17 October 2023 and 2 December 2024. Chinese labs had to train and serve models on fewer and weaker chips, which rewards designs that need less computing per token. DeepSeek trained its V3 model on a cut-down chip Nvidia made for the Chinese market.

The pressure eased in 2026. The US began allowing sales of more capable chips under conditions in January, and by May it had approved sales to about 10 Chinese companies, Reuters reported on 14 May 2026. Beijing has at times told its own companies to avoid some Nvidia chips, so supply remains tight.

Did DeepSeek really train a model for $5.6 million?

That figure covers only the final training run. DeepSeek's V3 paper (arXiv, December 2024) reports 2.788 million GPU hours for that run, which it valued at about $5.6 million. The paper says the figure excludes earlier research and experiments. The research firm SemiAnalysis estimated in January 2025 that DeepSeek had spent about $1.6 billion on servers.

The price you pay per answer mostly reflects the cost of serving it, plus each lab's pricing strategy, both covered above.

Did Chinese labs copy US models to cut costs?

Anthropic says some used its model. On 23 February 2026, it said DeepSeek, Moonshot and MiniMax had run more than 16 million exchanges with Claude through about 24,000 fraudulent accounts, to train their own models. By Anthropic's count, MiniMax accounted for over 13 million exchanges, Moonshot over 3.4 million and DeepSeek over 150,000. These are Anthropic's allegations.

Distillation means training a smaller model on a larger model's answers. Anthropic itself calls it "a widely used and legitimate training method", since labs routinely distill their own models; its objection is to competitors breaking its terms to copy Claude's abilities. Distillation can lower the cost of building a model. It does nothing to the computing each answer takes, which is what the per-token price pays for.

Is a cheaper token a cheaper answer?

Not always. A model that writes more tokens to reach an answer can cost more per task despite a lower price per token. On Artificial Analysis's numbers, GPT-6 Luna costs $0.07 per task against $0.27 for DeepSeek V4.1 Flash, at a similar score. GPT-6.1 Sol costs $0.72 per task for a score of 52, while Kimi K3 and GLM-5.3 cost about $2 for scores of 44 and 45. The standout Chinese results are both from Xiaomi: MiMo-V2.6-Pro scores 46 for $0.13 per task, and MiMo-V2.6-Flash scores 38 for $0.06, the cheapest per task in the table.

The US government's AI standards center found a similar pattern in 2025. A US reference model cost 35% less than the best DeepSeek model to reach similar performance across 13 benchmarks (NIST's Center for AI Standards and Innovation, 30 September 2025).

These tests are hard reasoning problems run at maximum effort. Everyday chat produces shorter answers, so the cost you'd see depends on how you use the model.

What this means if you're choosing an AI app

For everyday questions, I think a model slightly off the frontier is plenty. For the hardest coding and analysis, the US flagships still score highest, and how Chinese models compare in 2026 has the current scores. If you're worried about where your data goes, are Chinese AI models safe to use covers the apps, the hosts and the safety tests.

Low per-token prices, Chinese and now American, are what let a chat app charge a few dollars a month. Plainly costs $5 a month for unlimited chat. Plainly runs one capable open-source model and switches when a better one comes out. You never pick. Its past models include DeepSeek V4 Flash, DeepSeek V4.1 Flash, GLM 5.3 Flash and Xiaomi's v2.6 Pro, all from Chinese labs. The team sets out the numbers behind that price in how a $5 AI app makes money.