The honest framing: the headline price is real and remarkable. The full price — data stored in China, built-in censorship, self-hosting complexity — is what you weigh against it. For non-sensitive work the value is hard to beat; for regulated data, self-host the open weights or don't use them.
The major Chinese labs
| Lab | Model | Edge | Price /M |
|---|---|---|---|
| DeepSeek | V3 / R1 | MoE, budget coding value | $0.27/$1.10 |
| Alibaba | Qwen3 | 36T-token training, 1M context, multilingual | ~$0.38 |
| Moonshot AI | Kimi K2.6 | Agent-swarm; beat GPT-5.4 on SWE-Bench Pro | ~$0.60/$2.50 |
| Z.ai (Zhipu) | GLM-5.1 | 754B params, Huawei Ascend, MIT-licensed | ~$0.40/$1.50 |
| MiniMax | M3 | 80.5% SWE-bench as open weights | $0.30/$1.20 |
| Step | Step 3.5 Flash | 25x cheaper than GPT-4o, strong maths | $0.10/$0.30 |
| ByteDance | Doubao Seed 1.6 | Consumer + enterprise hybrid | Low |
| Baidu | ERNIE | Best Chinese-language, enterprise | Enterprise |
In one 12-day window in April 2026, four Chinese labs released open-weight coding models matching Western frontier performance at a fraction of the cost. Kimi K2.6 became the first open-weight model to beat GPT-5.4 on SWE-Bench Pro; Step 3.5 Flash ships at $0.10/$0.30 per million tokens.
Why they're this cheap
Two reasons. Mixture-of-Experts activates only the relevant slice of a giant model per query — the DeepSeek breakthrough. And US chip export controls (no H100/A100 since 2022) forced Chinese labs to innovate on software efficiency. The intended handicap became the competitive advantage.
The risk framework
Three risks decide whether a Chinese model fits your use case:
- Data sovereignty. Hosted APIs store data in China, subject to Chinese law. Multiple governments have restricted DeepSeek on these grounds.
- Content censorship. Models filter or refuse content on politically sensitive topics — usually irrelevant to business tasks, occasionally not.
- The mitigation. Self-hosting the open weights removes the jurisdiction risk entirely. The cheapest option is also the most complex to deploy safely.
Full detail in the China risk assessment and the data sovereignty comparison.
When to use them
| Good fit | Avoid (unless self-hosted) |
|---|---|
| Non-sensitive, high-volume tasks | Regulated personal / health / financial data |
| Budget coding and prototyping | GDPR-bound EU customer data via API |
| Self-hosted private deployment | Politically sensitive content work |
Considering the budget route? Model the savings against the work in the cheapest AI API guide and check where the data goes first.