Chinese LLMs via API: DeepSeek, Qwen and GLM with one key

DeepSeek, Qwen and GLM are the usual picks when you need an inexpensive model for high-volume tasks. Connecting each one directly is inconvenient: every vendor has its own sign-up, key and payment method. Through Zukko, these Chinese LLMs are available via API with one key and one balance. Below: a table of all seven models, a cost estimate on real volume, paying from Russia, and code for switching between models.

Chinese LLMs in the Zukko API: the table

Prices are in dollars per 1M tokens, as of October 6, 2026. The intelligence index is the Artificial Analysis Intelligence Index, an aggregate benchmark score.

These models have no separately charged cache write (the dash), and a cache read is counted if the provider reports it in the response.

Current models and prices: GET https://api.zukko.pro/v1/models or the Models page in the dashboard. The response format is described in the models docs↗.

How to read the table

The cheapest is deepseek-v4.1-flash. 0.013 for input and 0.038 for output. Its index (39) is also higher than that of the pricier DeepSeek V4 Pro (36). deepseek-v4-flash costs more than V4.1 Flash and has no index in the table — for a new project, by price and index, V4.1 Flash is the more logical starting point. Setup and calculations for it are in the article DeepSeek V4.1 Flash via API↗.

The highest index across the three lines is 45, shared by qwen3.8-max and glm-5.3. The difference: Qwen 3.8 Max has a 1M context and image input, GLM 5.3 has 128K and text only, but cheaper output (0.1 vs 0.263).

GLM charges the same for input and output. 0.1 / 0.1 for GLM 5.3 and 0.04 / 0.04 for Flash. That pays off where the answer is long relative to the request: text generation, rewriting, translation.

Image input is accepted only by qwen3.8-max and deepseek-v4-pro.

The 128K context applies only to GLM. For long documents and large codebases, use DeepSeek or Qwen with 1M.

Caching gives no discount. For all seven models a cache read costs the same as input, and a write is not charged separately. Don't count on a repeated prompt getting cheaper here.

The index is an aggregate score over general tests. For comparison: gpt-6.1-sol — 52, claude-opus-5-5 — 58. On your task the ranking may differ, so the final choice comes from testing on your own examples.

What it costs on real volume

Take a support bot that uses 10M input and 2M output tokens a month. Formula: cost = 10 × input price + 2 × output price.

The spread between the cheapest and the most expensive is about $1.32 a month at this volume. At these amounts the choice is usually decided not by price but by which model passes your quality check.

Paying from Russia

The Zukko balance is kept in dollars; top-ups start at $1. Options:

• SBP (Russia's fast payment system) and Russian bank cards (Mir, Visa, MasterCard) — paid in rubles at the market rate at the time of payment;
• cryptocurrency (USDT and others);
• a foreign card — with a commission of about 13%, amounts from $2 to $594.

There's no subscription: you pay only for tokens. There's no sign-up bonus either, but subscribing to the @zukkopro Telegram channel credits $0.50 — once per Telegram account: link Telegram in the dashboard, subscribe and press "Check". A step-by-step SBP guide is in the article on paying via SBP↗.

Switching between models with one key

All models are available at one address in the OpenAI format; switching means changing the model string. A Python example: the same question to three models, with token usage.

import os

from openai import OpenAI

client = OpenAI(base_url="https://api.zukko.pro/v1", api_key=os.environ["OPENAI_API_KEY"])

question = "Explain in two sentences what an idempotent request is."

for model in ["deepseek-v4.1-flash", "qwen3.8-flash", "glm-5.3-flash"]:
    r = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": question}],
    )
    print(f"--- {model}: input {r.usage.prompt_tokens}, output {r.usage.completion_tokens}")
    print(r.choices[0].message.content)

The same with curl:

curl https://api.zukko.pro/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "glm-5.3-flash", "messages": [{"role": "user", "content": "Hello"}]}'

If a model is temporarily off, the gateway returns 503 with the model_disabled code. A simple fallback chain:

import openai

CHAIN = ["deepseek-v4.1-flash", "qwen3.8-flash", "glm-5.3-flash"]


def ask(messages):
    for model in CHAIN:
        try:
            return client.chat.completions.create(model=model, messages=messages)
        except openai.APIStatusError as e:
            if e.status_code in (502, 503):
                continue
            raise
    raise RuntimeError("all models in the chain are unavailable")

Mind the context: if a request is longer than 128K tokens, GLM in the chain won't accept it — use it only for short requests.

In tools built on the OpenAI SDK (Cursor, Cline, LangChain and the like), set the base URL to https://api.zukko.pro/v1 and use your fetch_... key. Any of these models can also be called in the Anthropic format — the gateway converts the request. Parameters are in the chat/completions docs↗.

Choosing for your task

1. Filter by requirements: need images — Qwen 3.8 Max or DeepSeek V4 Pro; documents over 128K — DeepSeek or Qwen.
2. From what's left, take two or three by price and index.
3. Run them on 30–50 of your own examples with a clear "pass / fail" criterion.
4. Pick the cheapest one that passed. If none did, compare with gpt-6.1-sol or claude-sonnet-5 on the same examples.

Create a key under API keys↗; all seven models are available on it right away.

FAQ

Do I need to sign up with each vendor?

No. A Zukko account and key are enough: DeepSeek, Qwen and GLM, as well as Claude, GPT and Gemini, all work on one key with a shared balance.

Can I send an image to GLM?

According to the table, no: among these models only qwen3.8-max and deepseek-v4-pro accept image input.

Does prompt caching work for these models?

Their cache read price equals the input price and a write is not charged separately, so caching doesn't affect cost. Count all input at the input price.

Can I use them in Claude Code or Cursor?

Yes, any model can be called in both the OpenAI and Anthropic formats. Check how the tool works with your chosen model on your own project, and keep GLM's 128K context in mind.