The three Chinese flagships sell different things: DeepSeek sells throughput, Qwen sells breadth, Kimi sells long-context coding. All three cost a fraction of GPT-6 Astra's $10 per million input tokens.
Prices below are as listed on OpenRouter on 10 November 2026.
| Model | Lab | Input / MTok | Output / MTok | Context |
|---|---|---|---|---|
| DeepSeek V4 Pro | DeepSeek | $0.96 | $1.91 | 1,048,576 |
| Qwen3.8 Max | Alibaba | $2.00 | $6.00 | 1,000,000 |
| Kimi K3.1 Pro | Moonshot | $2.00 | $10.00 | 1,048,576 |
DeepSeek is the price
V4 Pro undercuts the other two on both sides of the meter. Its V4.1 Flash tier is listed at $0.01 input and $0.50 output, about a hundredth of its own flagship's input price — a number that reads like a promotion and should be checked before you build a budget on it.
If your workload is high volume and the answers are checked, this is the row that changes the arithmetic.
Qwen is the middle
At $2 and $6, Qwen3.8 Max has the cheapest output of the three at the top end. Alibaba also sells a $4 and $12 Max Prime tier, so the family spans a wide range under one name. Check which tier a provider is serving.
Kimi is the long-context coder
Kimi K3.1 Pro matches Qwen's $2 input and doubles its output at $10. Moonshot's pitch is coding over long contexts, and it is the most expensive of the three per answer. That is a deliberate position, not a mistake.
What none of this tells you
Context windows of one million tokens are now standard at all three labs and at OpenAI, so the number has stopped separating anyone. The difference is what each model does with the tokens, and that is what your own evaluation has to measure.
One more caveat: prices on a router are not always the lab's own price. DeepSeek publishes its API pricing directly; check there before committing to a monthly figure.