Guides · dev · deepseek api pricing
Chinese AI API prices, live, versus the latest Opus and Sol
The useful question is not “are Chinese models cheap?” Some are. A few flagships are only cheap if you accept a reseller’s rate, and one of them is expensive on output even at the lab’s own price.
This page is the map. The live table is the full list. The percentages below are rendered from the snapshot that built the site.
The flash tier is the real discount
Three general models sit together on the OpenRouter blend: Qwen3.8 Flash, GLM-5.3 Flash, and Hunyuan HY3. The chart states the gap against the current Opus and Sol rows. GLM-5.3 Flash is the one we also checked on the lab’s own page: Z.AI lists $0.15 input and $0.50 output per million tokens, which is the same pair as the OpenRouter card.
| Model | Blended | vs Opus | vs Sol |
|---|---|---|---|
| qwen3.8-flash | $0.230 | 97% cheaper | 94% cheaper |
| glm-5.3-flash | $0.291 | 96% cheaper | 93% cheaper |
| hy3 | $0.231 | 97% cheaper | 94% cheaper |
| deepseek-v4.1-flash | $0.198 | 98% cheaper | 95% cheaper |
| minimax-m3 | $0.525 | 93% cheaper | 87% cheaper |
DeepSeek’s own page is more precise than a single OpenRouter number. The API name deepseek-flash bills as V4.1 Flash. Peak is $0.30 input (cache miss) and $1.20 output. Off-peak is half of that. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays. The OpenRouter card matches the peak cache-miss rate, not the off-peak rate. If you can wait, the direct API is the cheaper one.
We have not re-checked a Qwen or Hunyuan invoice page for those flash ids. Until that is done, call them OpenRouter prices.
Flagships are a different market
Kimi K3, GLM-5.3, Qwen’s max tier, and DeepSeek V4 Pro do not belong in the flash sentence. Kimi’s output price dominates the blend. Qwen’s max tier can cost more than the Sol comparator on the same blend. The scatter makes the shape obvious: input can be pennies while output is double-digit dollars.
| Model | Input | Output |
|---|---|---|
| qwen3.8-flash | $0.150 | $0.470 |
| glm-5.3-flash | $0.013 | $1.13 |
| deepseek-v4.1-flash | $0.036 | $0.684 |
| minimax-m3 | $0.300 | $1.20 |
| glm-5.3 | $0.079 | $4.80 |
| kimi-k3 | $0.304 | $11.59 |
| qwen3.8-max-prime | $4.00 | $12.00 |
| deepseek-v4-pro | $0.955 | $1.91 |
| claude-opus-5.5 | $4.00 | $20.00 |
| gpt-6.1-sol | $2.00 | $10.00 |
Do not quote the GLM-5.3 or Kimi K3 list price alone
Z.AI’s own host on the OpenRouter endpoints feed matches the docs at $1.40 / $4.40. Moonshot’s own host matches the docs at $3 / $15, with cached input at $0.30 and a separate cache-write fee. The model-level list price is a different host. A blog that prints only the list price is describing a reseller.
What the percentage hides
A 75/25 blend assumes you send three tokens in for every token out. A coding agent that re-sends a long prompt and writes a short diff looks like that. A model that rambles, or a chain that retries, does not. The calculator uses an illustrative session instead of the blend. Cache-read prices matter there, and cache-write fees are still left out.
Small translation models (the Hunyuan MT2 rows) win the sort and lose the job. Sort the table by lab before you treat the cheapest row as a chatbot.
Where to go next
Questions
How is the cheaper-than percentage calculated?
Blended price is 75 percent of the input price plus 25 percent of the output price, per million tokens. That blend is divided into the same blend for the latest Opus and the latest Sol. The names of those two models come from the snapshot, not from this sentence.
Is the OpenRouter price the price the lab charges?
Sometimes. GLM-5.3 Flash matches Z.AI's page. GLM-5.3 and Kimi K3 do not. Read the endpoints note before you quote those two.
Sources
- OpenRouter models API, accessed 2026-10-16
- Z.AI pricing, accessed 2026-10-16
- DeepSeek API pricing, accessed 2026-10-16
- Kimi context caching, accessed 2026-10-16