Chinese LLM APIs really are cheaper, and four things to check before you switch
Across our 60-model dataset, the median Chinese-provider model is 2.1x cheaper on a blended basis than the median US or EU one. That gap is real. Whether you can use it is a different question.
Of the 17 providers we track, 7 are Chinese: DeepSeek, Moonshot AI, Zhipu AI, Alibaba, ByteDance, Baidu, StepFun. Between them they publish 14 of the 60 models in our dataset, and they are, as a group, conspicuously cheaper than everyone else.
That observation gets made often and quantified rarely. So here is the number, computed the same way for both groups.
The size of the gap
- Input: median $0.60 versus $1.38 — a 2.3x difference.
- Output: median $2.77 versus $5.00 — 1.8x.
- Blended (3:1): $1.13 versus $2.38 — 2.1x.
Medians rather than means, deliberately: a handful of very expensive frontier models would drag a Western average upward and overstate the gap. The median is the honest comparison and it still shows a substantial difference.
The cheapest Chinese-provider model in our dataset on a blended basis is Qwen3.7 Flash at $0.03 input and $0.13 output.
Why the gap exists
Several forces push in the same direction, and none of them is "the models are worse."
A genuinely more competitive domestic market
There are more serious model providers competing for Chinese developers than there are for Western ones, and they have been competing on price openly and aggressively. Price competition at that intensity produces price competition, which is not a tautology so much as an observation that Western providers have mostly not been in that fight.
Lower cost structures
Compute, power, and engineering labour cost less. Some providers are also part of larger technology groups with their own infrastructure, which changes what "cost" even means internally.
Different strategic goals
For several of these companies, the API is not the profit centre. Adoption, ecosystem position, and demonstrating capability matter more than per-token margin right now. Pricing reflects that priority, and it may not persist once the priority changes.
Efficiency work that shows up in the price
Some of these providers have shipped genuinely notable architectural and training-efficiency work. Cheaper inference per unit of capability is a real engineering result, not only a pricing decision, and it is reasonable to expect some of that gap to be durable.
Four things to check before switching
The gap is real. That does not make it automatically available to you. In rough order of how often it turns out to be the blocker:
1. Data residency and legal posture
This is the one that decides it for most regulated teams. Where does the data go, which jurisdiction's law governs it, and what does your customer contract or your regulator permit? If you have a data processing agreement that enumerates sub-processors, or you handle health, financial, or EU personal data, this question comes first and often ends the conversation regardless of price. Some providers offer non-mainland endpoints; check specifically rather than assuming.
2. Capability on your actual task
Benchmarks generalise badly. The only evaluation that matters is your own, on your own inputs, with your own definition of a good answer. Several of these models are genuinely excellent — and some are excellent at reasoning and code while being weaker at, say, low-resource languages or a specific domain vocabulary. Run a real eval set before migrating. A model that is 2.1x cheaper and fails 5% more often may cost more once retries and human correction are counted.
3. Latency and reachability from your region
A cheap token that arrives 800ms later is not cheap for an interactive product. Measure time-to-first-token and total latency from your actual serving region, not from your laptop. Check whether the provider has regional endpoints or whether every request crosses a continent. For batch and asynchronous work this barely matters; for a chat interface it can be decisive.
4. Operational maturity
Documentation quality, English-language support, SDK maintenance, status page transparency, deprecation notice periods, rate limit clarity, and billing predictability all vary. Some of these providers are excellent on all of it. Evaluate it as you would any vendor, because the cost of an unannounced breaking change is not measured in tokens.
The sensible way to use the gap
Framing this as an either/or is the mistake. Most production systems make several different kinds of model call, and they do not all carry the same constraints.
Classification, extraction, routing, translation, tagging, and internal summarisation are frequently high-volume, low-stakes, latency-tolerant, and — crucially — often operate on data that is already internal or already anonymised. That is precisely the profile where a cheaper model is an easy win. The user-facing call that needs your best answer, on your most sensitive data, can stay where it is.
Segmenting by call type rather than switching wholesale captures most of the saving while leaving your risk profile largely intact. It is also reversible, which a full migration is not.
The short version
- Median blended rate: $1.13 (Chinese providers) versus $2.38 (US/EU) — a 2.1x gap.
- It comes from a more competitive domestic market, lower cost structures, strategic pricing, and real efficiency work.
- Check residency, task-specific capability, latency from your region, and operational maturity — in that order.
- Segment by call type instead of migrating wholesale. Most of the saving, most of the risk avoided, and reversible.
Every rate above is in the comparison table, and each provider has its own page under the provider directory with its full model line and a link to its official pricing documentation.
How this was produced. Every price in this article is read at build time from the dataset behind the comparison table — 60 models across 17 providers, each rate taken from the provider's own published pricing page. Nothing is quoted from a secondary source or a third-party aggregator. The methodology page sets out how rates are checked and what is deliberately excluded, and the underlying numbers are downloadable as JSON and CSV if you want to redo any of the arithmetic yourself.
Analysis and judgements are the author's own. Published list prices change without notice; confirm against the provider before making a billing commitment.