Chinese LLM API Pricing in 2026: One Table for Six Providers
DeepSeek, Qwen, GLM, Kimi, MiniMax and MiMo are usually compared one at a time, in six different articles with six different currencies. This page puts the published per-million-token rates side by side, then explains the three things that actually change the invoice: peak windows, the CNY/USD gap, and the context tier you pick.
- Two providers publish in USD (DeepSeek, and gateways), four publish in CNY (GLM, Kimi, Qwen, MiMo, MiniMax) — the exchange rate is part of your price.
- Only DeepSeek has a peak surcharge (2x). Everything else is flat across the day.
- The cheapest line on paper is often not the cheapest line you can actually buy: mainland phone number and CNY payment are the usual blockers.
- Long context is a price tier, not a feature. Kimi K3 Pro 1M costs 4x the 128K tier on output.
Published rates per million tokens
| Provider | Model / tier | Context | Input / 1M | Output / 1M | Peak 2x? |
|---|---|---|---|---|---|
| DeepSeek | V4.1 Flash (off-peak) | long | $0.15 | $0.60 | Yes |
| DeepSeek | V4.1 Flash (peak) | long | $0.30 | $1.20 | — |
| DeepSeek | V4 Pro (off-peak) | long | $0.66 | $1.98 | Yes |
| Zhipu | GLM-5.2 Flash | 128K | ¥0.60 (≈$0.08) | ¥2.00 (≈$0.28) | No |
| Zhipu | GLM-5.2 Pro | 128K | ¥5.00 (≈$0.70) | ¥14.00 (≈$1.95) | No |
| Zhipu | GLM-5.2 Plus | 128K | ¥10.00 (≈$1.40) | ¥30.00 (≈$4.18) | No |
| Moonshot | Kimi K3 Flash | 128K | ¥0.60 (≈$0.08) | ¥2.00 (≈$0.28) | No |
| Moonshot | Kimi K3 Pro | 128K | ¥4.00 (≈$0.56) | ¥12.00 (≈$1.67) | No |
| Moonshot | Kimi K3 Pro 1M | 1M | ¥8.00 (≈$1.11) | ¥24.00 (≈$3.34) | No |
| Alibaba | Qwen Max | long | ¥2.00 (≈$0.28) | ¥6.00 (≈$0.84) | No |
| Xiaomi | MiMo v2.5 | long | from $0.14 (cache miss) | from $0.28 | No |
| MiniMax | M2 | long | from $0.15 | from $0.60 | No |
Rates as verified against each provider's pricing page on 2026-10-02 (DeepSeek), 2026-09-10 (GLM, Kimi) and the current text of our provider pages (Qwen, MiMo, MiniMax). CNY → USD conversions use ≈¥7.2/$. List prices change; check the provider page before budgeting. AI24X is an independent gateway and is not affiliated with any provider listed.
Three things that move the bill more than the sticker price
1. Peak windows exist on one provider only
DeepSeek charges 2x during 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. A workload that costs $0.60/M output off-peak costs $1.20/M at peak — the same tokens, twice the bill. GLM, Kimi, Qwen, MiMo and MiniMax publish flat rates, so scheduling only matters on DeepSeek. If you cannot control when batch jobs run, budget at the peak rate, not the off-peak one.
2. CNY pricing is not the same as USD pricing
GLM-5.2 Flash lists at ¥0.60 in / ¥2.00 out per million tokens. At ≈¥7.2/$ that is about $0.08 / $0.28 — cheaper than DeepSeek Flash off-peak on both lines. But you only realise that price if you can pay in CNY and settle at a reasonable rate. Paying through an international gateway moves the price to USD and adds a tier margin; the ranking can flip for small volumes.
3. Context length is a price tier
Kimi K3 Pro is ¥4.00/¥12.00 per million at 128K and ¥8.00/¥24.00 at 1M — double the input and double the output for the right to send longer prompts. The trap is that long context also means more input tokens per request, so cost per call grows faster than the per-token rate. If your prompts fit in 128K, staying in the 128K tier is usually the single biggest saving available.
Worked example: 30M input + 6M output per month
The same monthly volume priced on each option (30 million input tokens, 6 million output tokens):
| Option | Input bill | Output bill | Monthly total |
|---|---|---|---|
| DeepSeek Flash, all off-peak | 30 × $0.15 = $4.50 | 6 × $0.60 = $3.60 | $8.10 |
| DeepSeek Flash, all peak | 30 × $0.30 = $9.00 | 6 × $1.20 = $7.20 | $16.20 |
| GLM-5.2 Flash (CNY) | 30 × ¥0.60 = ¥18.00 | 6 × ¥2.00 = ¥12.00 | ¥30.00 (≈$4.18) |
| Kimi K3 Flash (CNY) | 30 × ¥0.60 = ¥18.00 | 6 × ¥2.00 = ¥12.00 | ¥30.00 (≈$4.18) |
| Kimi K3 Pro 1M (CNY) | 30 × ¥8.00 = ¥240.00 | 6 × ¥24.00 = ¥144.00 | ¥384.00 (≈$53.48) |
| AI24X flash tier (flat output) | billed separately | ≈6 × $0.35 = $2.10 | varies by routing |
Two lessons fall out of the table. First, the 1M-context tier is roughly 13x the 128K tier for the same volume — it is a tool, not a default. Second, the difference between the best and worst option here is under $12/month at 36 million tokens, which means engineering time to switch providers is usually worth more than the price gap until you reach hundreds of millions of tokens.
Getting access from outside China
This is where most pilots stop. The providers' own consoles generally expect a mainland phone number for verification and a CNY payment method, and some require real-name verification. Two workable paths:
- Gateway access. One OpenAI-compatible key, billed in USD, often payable by international card or PayPal. You trade a tier rate for not needing a Chinese identity. See paying for Chinese LLM APIs with PayPal.
- Bring your own keys. If you already hold provider accounts, a BYOK gateway gives you routing, failover and usage reporting without reselling tokens. See BYOK AI gateway.
Choosing between them
| If your workload is… | Start with | Why |
|---|---|---|
| Everyday chat, extraction, light coding | DeepSeek V4.1 Flash or GLM-5.2 Flash | Lowest published output rates in the set; flat pricing on GLM |
| Long documents, whole-codebase context | Kimi K3 Pro 1M | The 1M tier is the differentiator; pay for it only when the prompt needs it |
| Multilingual, tool calling, balanced cost | Qwen Max | Flat CNY pricing, broad modality support through the Alibaba stack |
| Cheap classification at high volume | MiMo v2.5 | Low input rate, and cache hits are discounted heavily |
| Mixed traffic with unpredictable load | A gateway with routing | Moves each request to a tier that clears the quality bar, with failover |
Provider details
FAQ
What is the cheapest Chinese LLM API?
On published list prices, DeepSeek V4.1 Flash off-peak ($0.15 in / $0.60 out) and GLM-5.2 Flash (¥0.60 / ¥2.00, about $0.08 / $0.28) are the low end of the set. DeepSeek bills in USD, GLM and Kimi bill in CNY, so the winner depends on the exchange rate you actually settle at.
Can I use them without a Chinese phone number?
Not through the providers' own consoles in most cases — verification generally expects a mainland number. Gateways that hold their own provider accounts remove that requirement entirely.
Are Chinese LLM APIs cheaper than OpenAI?
At comparable tiers, yes. DeepSeek Flash off-peak lists at $0.15 / $0.60 per million tokens against $5.00 / $20.00 for GPT-5. Against OpenAI's smallest models the gap narrows to near zero.
Which one has the largest context window?
Kimi K3 Pro 1M at a one-million-token window (¥8.00 in / ¥24.00 out per million). Longer context is priced as a separate tier, so it is worth checking whether your prompts genuinely need it.
Do I need a Chinese bank account?
For the providers' consoles, in practice yes: Alipay, WeChat Pay or a China-issued UnionPay card. Gateways accept international cards and PayPal and bill in USD.
Educational content. Rates are list prices published by each provider and are shown for comparison only; they change frequently. AI24X is an independent gateway and is not affiliated with DeepSeek, Zhipu, Moonshot, Alibaba, Xiaomi or MiniMax.
Model pages: all supported models · DeepSeek · Qwen · Kimi · GLM · MiniMax · MiMo