Why are Chinese AI models so cheap?
In our updated market reports, you will find everything you need
SUMMARY
Chinese AI models are so cheap because Chinese laboratories have sharply reduced the compute and memory required for each token, while a brutal domestic price war has pushed customer prices even lower than those engineering gains alone would justify.
The label hides a wide market. DeepSeek and MiniMax can undercut premium US models by roughly 20 times on a simple blended-token comparison, while Kimi already charges much closer to Western frontier pricing.
The main structural advantage is sparse Mixture-of-Experts design. Models with hundreds of billions or even trillions of parameters may activate only about 3% to 5% of them for each token, preserving broad capability without paying the full compute bill every time.
Long context has become another major source of savings. Compressed attention, sparse attention, lower-precision arithmetic, fused kernels and cache-aware serving reduce the amount of data that must be stored and moved, which is especially valuable for coding agents and research tools handling enormous prompts.
US chip restrictions helped create the pressure for this efficiency, but they did not make Chinese hardware cheap. The restrictions raised the cost and complexity of compute; Chinese teams clawed back part of that handicap through communication compression, hardware-aware training and support for several accelerator stacks.
The price war then transferred much of the technical saving to customers. Cuts of up to 97%, free business access, launch discounts and large promotional quotas moved far faster than the underlying cost of chips or electricity, so margin sacrifice was plainly part of the strategy.
The financial accounts fit that picture. Independent model companies are posting losses far above revenue, while groups such as Alibaba can treat cheap model access as a way to sell cloud infrastructure, databases, storage and enterprise services around it.
Open weights intensify the pressure worldwide because the original laboratory no longer controls the only place the model can run. Competing hosts can optimize the same weights, while subsidies and distillation lower selected parts of the bill without explaining the full infrastructure advantage.
Cheap tokens do not always mean a cheap production workload. Longer reasoning traces, retries, agent loops, latency and human review can reverse the apparent saving, so the useful metric is cost per successful task rather than the posted price per million tokens.
China is now developing a two-tier AI market: near-commodity intelligence sold at razor-thin prices, and newer frontier capability that recovers a premium sooner. Better engineering made the low prices possible; the fight for developers, cloud workloads and market share pushed them to their current extremes.
Are Chinese AI models really that cheap right now?
Yes. At current API rates, several leading Chinese models cost far less than premium US models, although the gap ranges from about two times to more than 20 times depending on which models we compare.
DeepSeek V4 Pro currently charges $0.435 per million uncached input tokens and $0.87 per million output tokens. MiniMax M3 is even cheaper for contexts below 512,000 tokens, at $0.30 and $1.20. Z.ai's GLM-5.2 sits higher, while Moonshot's Kimi K3 has moved into a much more expensive tier. The Chinese market already spans a wide price range.
The comparison becomes striking at the premium end. OpenAI lists GPT-5.6 Sol at $5 per million input tokens and $30 per million output tokens. Anthropic lists Claude Opus 5 at $5 and $25. With a workload containing three input tokens for every output token, DeepSeek V4 Pro comes out around 21 times cheaper than GPT-5.6 Sol, while MiniMax M3 is around 19 times cheaper than Claude Opus 5. OpenAI's cheaper Luna model narrows the difference substantially.
These figures compare list prices before cache discounts, batch discounts, free quotas or negotiated enterprise rates. They also say nothing about quality, speed or how many tokens a model needs to finish the job. We will deal with those differences separately.
| Model | Input per 1M tokens | Output per 1M tokens | Blended price at 3:1 input/output |
|---|---|---|---|
| MiniMax M3, context up to 512K | $0.30 | $1.20 | $0.53 |
| DeepSeek V4 Pro | $0.435 | $0.87 | $0.54 |
| GLM-5.2 | $1.40 | $4.40 | $2.15 |
| OpenAI GPT-5.6 Luna | $1.00 | $6.00 | $2.25 |
| Kimi K3 | $3.00 | $15.00 | $6.00 |
| Claude Opus 5 | $5.00 | $25.00 | $10.00 |
| OpenAI GPT-5.6 Sol | $5.00 | $30.00 | $11.25 |
What does “cheap” mean for a Chinese AI model?
Here, “cheap” can mean four different things: a small final training bill, a low full research cost, inexpensive inference, or a low price for customers.
Those four numbers often move in opposite directions. One laboratory may train efficiently and charge a healthy margin. Another may pour money into research, then sell API access at a loss to win developers. A third may release the weights and let outside hosts compete away the serving margin.
DeepSeek's widely quoted V3 training figure illustrates the confusion. It covered the reported GPU time for the final training process, using an assumed rental rate for each H800 GPU-hour. The paper explicitly left out earlier architectural research, experiments, salaries, data work and the cost of acquiring and running the cluster.
The API price can be far below the laboratory's full cost because companies set it according to spare capacity, competition, cloud strategy and how badly they want adoption. The real investigation has two parts: how Chinese labs lowered the cost of running a model, and why their companies passed so much of the saving to customers.
Did China’s AI price war make these models artificially cheap?
Yes. The technical savings are real, but customer prices fell too far and too fast to be explained by engineering alone.
The break came when ByteDance, Alibaba, Baidu and Tencent began answering one another with abrupt cuts, free tiers and promotional quotas. Alibaba reduced some Qwen prices by as much as 97%, including a fall in Qwen-Long's input rate from CNY 0.02 to CNY 0.0005 per thousand tokens. Baidu made two Ernie models free for business users soon afterward.
A 97% reduction means the price fell by a factor of 33. Nothing in the hardware or electricity bill improved 33-fold over the same few weeks. Providers had room to give up margin, and some were plainly willing to subsidize usage to stop rivals gaining ground.
The same behavior is visible today. Alibaba's current domestic price sheet applies temporary discounts of 20% to 50% to several Qwen models, alongside cache reductions, batch discounts and free allowances. The company labels the lower rates as promotional rather than permanent.
The price war explains why customers captured so much of the engineering gain. Chinese labs found cheaper ways to train and serve models, and domestic competition forced them to hand much of the saving straight to users.
If you want more recent data on this point, please see our latest foundation model market report.
Are Chinese AI companies losing money to keep prices low?
Often, yes. The public accounts show that several Chinese model companies are spending far more than they earn, although those losses do not prove that every API request is priced below its immediate serving cost.
MiniMax reported $79 million in revenue for 2025 and an adjusted net loss of $250.9 million. Zhipu's first annual report showed CNY 724 million in revenue and a CNY 4.718 billion net loss. Both companies were funding research, infrastructure and expansion, so we cannot assign the entire deficit to cheap tokens. The scale of the losses still makes one point hard to dodge: these businesses are optimizing for position and growth rather than mature software margins.
The large technology groups can play an even longer game. Alibaba has committed CNY 380 billion over three years to cloud and AI infrastructure. Its cloud revenue is now growing around 40%, and the company expects AI model and application services to become a major share of external cloud revenue. A cheap Qwen call can pull in storage, databases, networking and enterprise workloads that earn money elsewhere.
Earlier cloud wars followed the same playbook: make the headline service irresistible, then earn money from the surrounding stack. Some Chinese API prices probably carry very little margin, and some promotions may lose money. Public disclosures do not let us calculate the loss on each token with confidence.
Why do giant Chinese models use so little compute per token?
Sparse Mixture-of-Experts architecture is currently the biggest technical reason: the model may contain hundreds of billions or trillions of parameters, yet it wakes up only a small share for each token.
Think of the model as a huge company with specialist teams. Every question is routed to a few relevant teams instead of calling the entire workforce into one meeting. The system keeps the breadth of a very large model while paying a much smaller computation bill for each answer.
DeepSeek V4 Pro contains 1.6 trillion parameters but activates 49 billion per token, roughly 3.1%. Its Flash version activates 13 billion out of 284 billion. MiniMax M3 uses about 23 billion out of 428 billion, while GLM-5.2 uses 40 billion out of 753 billion.
American laboratories use sparse models too. What stands out in China is the consistency: DeepSeek, MiniMax, Z.ai and Moonshot have all made low active-parameter counts central to their products. Expensive and constrained compute has clearly shaped their designs.
Sparse models create difficult routing, networking and load-balancing problems. Chinese labs have spent heavily on those less glamorous details. The low token cost comes from making the whole system work efficiently, not from displaying an impressive parameter ratio.
| Model | Total parameters | Active per token | Active share |
|---|---|---|---|
| DeepSeek V4 Pro | 1.6T | 49B | 3.1% |
| DeepSeek V4 Flash | 284B | 13B | 4.6% |
| GLM-5.2 | 753B | 40B | 5.3% |
| MiniMax M3 | 428B | 23B | 5.4% |
If you want more recent data on this point, please see our latest foundation model market report.
How did DeepSeek make a one-million-token context affordable?
DeepSeek made a one-million-token context affordable by shrinking what each new token must read and what the hardware must remember.
A model processing a huge document does more than perform arithmetic. It keeps a growing memory of previous tokens and repeatedly moves data through scarce high-bandwidth memory. At long context lengths, that movement can become the bottleneck even when the GPU still has spare mathematical capacity.
DeepSeek V4 combines compressed attention with sparse attention. At a one-million-token context, DeepSeek reports that V4 Pro uses 27% of the single-token computation and 10% of the key-value cache required by V3.2. V4 Flash falls to 10% of the computation and 7% of the cache. The company also uses FP4 quantization-aware training and fused kernels, reporting inference speedups of roughly 1.5 to 1.73 times in general serving.
Moonshot has attacked the same bill from the infrastructure side. Its Mooncake serving system separates prefill work from token generation and treats the key-value cache as a shared resource. In published tests, that design raised throughput by as much as 525% in simulation and handled 75% more requests under real workloads.
Cheaper electricity or labor cannot account for gains of this size. The laboratories are redesigning how models store context, move data and schedule requests, which pays off especially well for coding agents and research tools that repeatedly send enormous prompts.
Did chip sanctions force Chinese labs to become more efficient?
The sanctions sharpened the pressure to save compute, but they also made Chinese AI hardware more expensive and harder to obtain.
DeepSeek-V3 was trained on 2,048 Nvidia H800 GPUs, a China-specific product with restricted interconnect bandwidth. The team responded with communication compression, overlapping data transfers with computation and a hardware-aware training stack. DeepSeek's own follow-up paper describes the model as a case study in designing around limits in memory, computation and networking.
That pressure continues. DeepSeek says V4 kernels have been validated on both Nvidia GPUs and Huawei Ascend processors. Its paper also argues that careful overlap between communication and computation can reduce the value of ever-higher interconnect bandwidth once a minimum threshold is reached.
The cost of the restrictions remains substantial. Current reporting still places Nvidia's top accelerators well ahead of Huawei's best chips in per-chip performance, while Chinese companies must support several hardware stacks and cope with weaker access to advanced memory and manufacturing.
Scarcity encouraged better engineering. But saying sanctions made Chinese compute cheap gets the direction wrong: the restrictions raised the hardware handicap, and laboratories recovered only part of that loss through software.
If you want more recent data on this point, please see our latest foundation model market report.
Is DeepSeek’s famous $5.6 million training figure still useful?
Yes, but only for one narrow purpose: it measures a highly efficient final training run, not everything DeepSeek spent to become a frontier laboratory.
The V3 technical report recorded 2.788 million H800 GPU-hours for pretraining, context extension and post-training. Multiplying those hours by DeepSeek's assumed $2 rental rate produced $5.576 million. That figure is detailed enough to take seriously and remains remarkable for a model of V3's scale.
The same report excludes earlier research, architectural experiments and data work, along with the broader cost of buying and operating the GPU cluster. A new company with that budget could not reproduce DeepSeek-V3 from scratch.
The figure remains useful because it measures the final recipe after the laboratory had already learned what to build. That repeatable part of the bill can fall quickly as the recipe improves. It also helps explain low API prices: once research and infrastructure are already paid for, the next training run and each additional request can cost far less than outsiders assumed.
Are Chinese government subsidies paying for the low prices?
Subsidies pay part of the bill, especially for compute access and early adoption, but they cannot explain the engineering gap on their own.
Shanghai's current AI support program gives the scale of the intervention. The city allocated CNY 600 million for compute vouchers, with combined city and district support able to cover up to 100% of eligible rental costs for a limited period. It also set aside CNY 300 million for model vouchers that can cover up to half the cost of third-party model APIs or private deployments, subject to caps.
Programs like these help startups use models sooner, keep local data centers busy and soften the cash burden of experimentation. National and local industrial funds also support chipmakers, infrastructure and research. RAND's review of Chinese AI policy identifies subsidized compute and application deployment as important tools in the country's strategy.
Vouchers can pay for compute, but they cannot invent sparse attention, compress a key-value cache or stabilize low-precision training. DeepSeek, MiniMax, Z.ai and Moonshot publish different systems built by separate teams. Public support gives them more room to experiment; engineers still have to create the reduction in compute per token.
Is China saving money by copying American models?
Distillation probably saves Chinese laboratories some post-training work, but the public evidence does not support the idea that copying explains why their models are so cheap.
Distillation is common across the industry. A smaller or newer model studies answers produced by a stronger model, which can replace part of the human labeling and reinforcement-learning work. DeepSeek openly distilled its own R1 reasoning into smaller Qwen and Llama models. V4 also uses several internal teacher models to combine specialized skills into one system.
Anthropic currently says it has detected large-scale attempts by Chinese groups to extract knowledge from Claude and argues that such activity can help foreign laboratories close capability gaps. Those are serious allegations from a company with access to its own traffic logs. Anthropic has not published enough case-level evidence for outsiders to calculate how much any named Chinese model gained from unauthorized distillation.
Much of the cost advantage is visible in original technical work: sparse expert routing, compressed attention, FP4 and FP8 training, cache-aware serving and custom communication systems. Copying outputs could speed up behavior tuning, especially in coding and reasoning, but it cannot produce those infrastructure gains by itself.
We can say confidently that distillation is part of modern model development. Public information does not tell us how much any Chinese laboratory gained from unauthorized use.
If you want more recent data on this point, please see our latest foundation model market report.
Why do open weights drive Chinese AI prices down worldwide?
Open weights remove the original laboratory's control over hosting, so several providers can compete to run the same Chinese model.
DeepSeek V4, MiniMax M3 and Kimi K3 are currently available as downloadable weights. A business can host them itself, hire a specialist inference provider or switch between competing hosts without retraining the model from zero.
The economics change immediately. A closed model leaves the developer with the owner's API price. Downloadable weights let hosting companies compete on utilization, quantization, speed and margin, while large customers can run the model on their own machines and keep sensitive data in-house.
The laboratory may earn less direct API revenue as a result. Chinese companies often accept that trade because open distribution builds developer mindshare, encourages compatible tools and can pull workloads toward their clouds or commercial services. It also makes a domestic model harder to displace: thousands of outside teams begin optimizing and adapting it.
Open weights reduce prices around the world without changing what the original training run cost. One model becomes a small market of competing implementations.
Are Chinese models cheap because they are weaker?
Sometimes, but the current quality gap is generally much smaller than the largest price gap.
Artificial Analysis currently gives Claude Opus 5 an intelligence score of 61 and GPT-5.6 Sol a score of 59. Kimi K3 reaches 57, GLM-5.2 reaches 51 and DeepSeek V4 Pro reaches 44. The same evaluation estimates task costs of roughly $2.03 for Opus 5, $1.54 for GPT-5.6 Sol, $0.72 for Kimi K3, $0.32 for GLM-5.2 and $0.04 for DeepSeek V4 Pro.
The models remain far from interchangeable. Premium American systems lead on many difficult scientific, coding and agent tasks, and benchmark averages can hide reliability problems. DeepSeek's 44 is a meaningful distance from 59 or 61. A company may gladly pay more when one failed legal analysis, broken deployment or missed security issue costs far more than the API bill.
Even so, a measured DeepSeek task costing about one-fortieth as much as GPT-5.6 Sol does not deliver only one-fortieth of the useful work. For summarization, extraction, translation, routine coding and large fleets of subagents, “good enough at huge volume” can beat a small quality advantage.
Kimi K3 stops us from treating every Chinese model as a bargain. At $3 per million input tokens and $15 per million output tokens, it sits much closer to premium Western pricing than DeepSeek or MiniMax. China currently offers several exceptional price-to-performance options, not one uniformly cheap national category.
If you want more recent data on this point, please see our latest foundation model market report.
Why can a cheaper API still cost more in production?
A lower token rate can produce a higher final bill when the model thinks for longer, repeats work or fails often enough to require another run.
Reasoning models vary enormously in output length. A recent study found that the model with the lower advertised rate became more expensive in 21.8% of pairwise comparisons once actual reasoning-token use was counted. The worst reversal reached 28 times, and repeated runs of the same prompt could vary sharply in length.
Agentic coding widens the problem. Another study found that coding agents can consume around 1,000 times more tokens than ordinary code chat, with a 30-fold difference between systems attempting the same task. Extra tokens did not reliably produce higher accuracy.
Language and tokenization can change the count too, usually by much less than a long reasoning trace or a failed attempt. The number worth tracking is cost per successful task, including retries, latency and human review. A model that is perfect for bulk extraction may become an expensive choice for an autonomous coding workflow.
Companies should test models on their own prompts and record complete-task costs. The list price is useful, but it is only the first line of the invoice.
Are Chinese AI prices already starting to rise?
Yes, at the frontier end. China's cheapest APIs remain extremely inexpensive, while newer premium models are beginning to charge more and limit subsidized access.
As seen above, Kimi K3 currently costs $3 per million input tokens and $15 per million output tokens, placing it six to 11 times above DeepSeek V4 Pro on a simple blended comparison. Demand was strong enough that Moonshot temporarily paused new subscriptions while it added capacity.
Z.ai has also moved customers away from older plans after a recent price increase. Alibaba's newest Qwen price sheet separates original rates from temporary launch discounts instead of promising that the promotional price will last. ByteDance has tested paid Doubao tiers reaching CNY 500 per month.
The market is learning to charge different customers differently. Cheap flash models handle high-volume routine work, while frontier reasoning, huge context windows, priority capacity and agent products carry higher prices. Providers can keep an attractive entry rate and monetize speed, reliability, subscriptions or surrounding cloud services.
The bargain era is changing, not disappearing. Commodity intelligence keeps getting cheaper, while the newest capability recovers a premium sooner than it did during the first phase of the price war.
Can American AI companies erase the price gap?
American providers can close much of the gap in economy models, but Chinese competition is likely to keep pressure on margins and accelerate every price cut.
OpenAI already spans a wide range, from GPT-5.6 Sol at $5 and $30 per million tokens to Luna at $1 and $6. Anthropic offers caching and batch discounts that can reduce effective costs sharply. Google, Meta, Nvidia and independent hosts compete lower in the market with smaller or open-weight systems.
The technical ideas also travel. Sparse experts, low-precision arithmetic, cache compression, speculative decoding and model distillation are published, copied and improved across borders. A Chinese laboratory may lead one generation without keeping the method exclusive.
Long-run data supports continued price compression. A recent study tracking hundreds of models estimated that economy-tier token prices fell about 600-fold from 2020 to 2026 and have recently been halving roughly every 1.1 years. It also found a break around the start of China's model price war, when competitive pressure began driving prices down faster. Premium reasoning prices proved much stickier.
That pattern is likely to persist. US companies will offer cheap models wherever intelligence has become interchangeable, while charging more for the newest capability, stronger tools, enterprise support and trusted deployment until competitors narrow those gaps too.
Why are Chinese AI models so cheap?
Mostly because Chinese laboratories have cut the compute and memory used for each token, while Chinese companies charge even less aggressively than those technical gains alone would require.
The engineering comes first. Sparse models activate only a small share of their parameters per token. Compressed and sparse attention reduce the cost of huge contexts. Lower-precision training, custom kernels and better scheduling squeeze more work from constrained hardware. These are measurable improvements described across DeepSeek, MiniMax, Moonshot and Z.ai, so the explanation reaches well beyond one headline model.
The commercial battle then amplifies the saving. Alibaba's 97% cuts, free tiers, temporary discounts and the heavy losses reported by independent model companies show that providers have been buying adoption. Large groups can recover value through cloud infrastructure and consumer products. Startups can accept losses while investors and public markets fund the race for scale.
Open weights add another layer of price pressure by letting several hosts compete over the same model. Government vouchers lower compute and deployment costs for some users. Huge domestic traffic improves utilization. Distillation can make post-training cheaper. Each factor helps, but none matches the combined effect of sparse engineering and a brutal price war.
One thing has changed lately: the cheap label no longer fits every major Chinese model. DeepSeek and MiniMax still set extraordinarily low rates, while Kimi K3 and other frontier offerings are moving toward premium pricing. China is creating a two-tier market, with near-commodity intelligence sold at razor-thin prices and newer capability monetized before it too becomes cheap.
Our final judgment is clear. Chinese AI models are structurally cheaper to run than many outsiders expected, and current prices sit below what engineering alone would predict. Better systems made those prices possible; the fight for developers, cloud workloads and market share pushed them to today's extremes.
If you want more recent data on this point, please see our latest foundation model market report.
OUR METHODOLOGY
This analysis separates four meanings of “cheap”: the cost of the final training run, the laboratory's full research cost, the operating cost of inference, and the price charged to customers. Keeping those numbers separate prevents a low API rate or a narrow GPU-hour estimate from being mistaken for the full economics of a frontier model.
We used official API price sheets to compare customer costs, technical reports and repositories to examine architecture, memory use and serving efficiency, company disclosures to understand investment priorities and losses, and independent research for broader quality and market comparisons. Each source was used only for what it could directly establish.
We gave the most weight to explanations that appeared across several laboratories and several evidence types. Sparse expert routing, compressed attention, low-precision training, cache-aware serving and aggressive price competition were supported by recent, measurable evidence. Sanctions, subsidies and distillation were treated as narrower influences on specific parts of the cost structure rather than complete explanations.
Key sources include DeepSeek's API pricing, MiniMax's pay-as-you-go pricing, OpenAI's API pricing, Anthropic's pricing, Z.ai's model pricing, Moonshot's Kimi pricing, DeepSeek's V4 release materials, the DeepSeek-V3 technical report, DeepSeek's hardware-aware infrastructure analysis, and Moonshot's Mooncake serving paper.
Additional company and policy context came from Alibaba's CNY 380 billion AI and cloud investment announcement, Alibaba's fiscal-year 2026 results, Anthropic's report on distillation attacks, Shanghai's AI support programs, RAND's review of Chinese AI policy, Artificial Analysis model comparisons, and the public model and implementation materials hosted by Hugging Face and GitHub.
