DEV Community

Cover image for Chinese LLM API pricing in 2026: 114 models and an 85x spread
HAOHAOnovai
HAOHAOnovai

Posted on

Chinese LLM API pricing in 2026: 114 models and an 85x spread

If you are choosing a Chinese model right now, the honest description of the problem is this: the cheapest and the most expensive option, reachable through the same OpenAI-compatible endpoint and billed in the same unit, differ by 85x.

$0.17 versus $14.50 per million output tokens.

That is not a rounding difference you can ignore while you "just ship something". It is the difference between a feature that costs you $200/month and one that costs you $17,000/month for the same token volume.

So I pulled the numbers. All of them are measured from a single GET /v1/models response plus the published price table, read on 2026-11-17. Nothing here is copied from a vendor's landing page.

The shape of the catalogue

Models reachable through one API key 114
Priced per token 57
— with an output price (chat / generation) 51
— embeddings (input price only) 6
Billed per generation (video / image / 3D / speech / music) 59
Output price range $0.17 → $14.50 per 1M = 85x
Input price range $0.043 → $2.90 per 1M

Cheapest 10 by output price

Model Provider Input $/1M Output $/1M
deepseek-v4-flash DeepSeek 0.080 0.170
hy-mt2-lite Tencent Hunyuan 0.043 0.172
glm-5.3-flash ZhiPu 0.080 0.280
mimo-v2.6-flash Xiaomi 0.100 0.280
hy-mt2-plus Tencent Hunyuan 0.072 0.287
glm-5.3-flashx ZhiPu 0.120 0.400
doubao-seed-2.0-lite ByteDance 0.075 0.450
qwen-plus Alibaba Cloud 0.185 0.530
qwen3-omni-flash Alibaba Cloud 0.150 0.600
minimax-m3.1-flash MiniMax 0.150 0.600

Most expensive 6 by output price

Model Provider Input $/1M Output $/1M
kimi-k3 Moonshot 2.900 14.500
kimi-k3.1-pro Moonshot 2.000 10.000
kimi-k2.7-code-highspeed Moonshot 2.073 8.610
qwen3.8-max Alibaba Cloud 1.840 5.515
doubao-seed-evolving ByteDance 0.920 4.620
glm-5.3 ZhiPu 1.250 4.000

The cheap end is spread across four providers. The top three all come from one. That matters for failure modes as much as for price: a single provider's outage is a bigger slice of your catalogue than you think.

The unit trap: 57, 51, 6, 59

This is where most price comparisons quietly go wrong.

57 models are priced per token. But 6 of those are embeddings — they have an input price and no output price at all, because they return a vector, not a completion. Comparing an embedding model's price to a chat model's output price is a category error.

59 models are billed per generation — video, image, 3D, speech, music. Those are quoted per clip or per image. Put them in a per-token table and the resulting number means nothing.

And 114 is the count of reachable models, while 57 is the count of per-token ones. Those are answers to different questions, and mixing them produces confident nonsense.

Turning prices into cost

List price is the starting point. Three things move the real number, roughly in order:

  1. Retries. A model needing 1.3 attempts per task costs 30% more than its list price. On agent workloads this dominates every difference in the tables above.
  2. Cache discounts. Cached input is often billed at a fraction of list. A long, stable system prompt can change the effective price substantially — and asymmetrically between providers.
  3. Output length. Output typically costs 3–20x the input price. Two models with identical input prices can differ 2x on real cost depending on how verbose they are on your task.

Which is why the only comparison that survives contact with production is cost per completed task on your own workload — not price per million tokens.

What this page deliberately does not tell you

It contains no latency data. Latency is the half of cost that decides whether a model is usable at all, and it cannot be read off a price table. Measuring it honestly needs controlled conditions: same region, same concurrency, same input length, repeated at different times of day. That is a measurement programme, not an afternoon of reading price tables — and a comparison that skips it and calls the result "cost" is telling half the story.

How the numbers were produced

  • Model count: GET /v1/models, 114 entries, read 2026-11-17.
  • Prices: the published per-model table, normalised to USD per 1M tokens.
  • Unit classification: "per token" if input or output price is non-zero; "per generation" if both are zero.
  • The table is refreshed weekly; the snapshot date is the authoritative "as of".

If a number here is wrong, the correct fix is a corrected table — not a rephrased claim. The full, machine-readable index is at aiapi-pro.com/chinese-ai-api-pricing, and the same catalogue is served by the novai-prices-mcp MCP server if you would rather query it from an agent than read a page.


If you want the same table as data rather than prose — the index page has it, filterable by billing unit.

Top comments (0)