I wanted a smaller OpenRouter catalog for Chinese models, so I built one

#ai #api #python #machinelearning

OpenRouter is useful. I still use it when I need broad model coverage.

But for one part of my work, that broad catalog became friction. I mainly wanted the current flagship models from Qwen, Kimi, GLM, MiniMax, and DeepSeek. I kept sorting through old versions, free routes, provider variants, and model IDs.

So I built Vancine, a smaller OpenAI-compatible API focused on current Chinese models.

Disclosure: I run Vancine. This is a product post, not an independent comparison. English is not my first language, and I used an AI writing assistant to edit the wording.

What it does

Vancine uses one API key, one balance, and one OpenAI-compatible endpoint:

https://vancine.com/v1

The catalog is intentionally small. When a Chinese provider replaces a flagship model, I update the catalog and retire the older listing instead of keeping every generation around.

The same account also reaches Chinese image, video, audio, and 3D models. I started with text and coding workloads, but media access has become useful for people who do not want separate accounts with several Chinese providers.

The price comparison

These are the four exact paid listings I compared on August 27, 2026. Prices are USD per 1M tokens.

Model	Vancine input / output	OpenRouter input / output
qwen3.8-max	$1.60 / $4.80	$2.00 / $6.00
kimi-k3	$2.40 / $12.00	$3.00 / $15.00
glm-5.3	$1.12 / $3.52	$1.40 / $4.40
MiniMax-M3	$0.24 / $0.96	$0.30 / $1.20

That is 20% lower for these four listings. Free variants, promotional routes, and temporary provider discounts are excluded. I am not claiming that every model is cheaper.

The live pricing page is the source of truth because provider prices can change.

Moving an existing OpenAI client

The change is small:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["VANCINE_API_KEY"],
    base_url="https://vancine.com/v1",
)

response = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[
        {"role": "user", "content": "Explain this failing test."}
    ],
)

print(response.choices[0].message.content)

OpenRouter model IDs include a provider prefix. Vancine IDs do not, so qwen/qwen3.8-max becomes qwen3.8-max.

Chat completions and streaming follow the OpenAI-compatible format. Provider-specific errors may differ, so I would still test the exact tool and model combination before moving production traffic.

What I have actually tested

I have run Pi against Kimi K3 through Vancine on a real edit-and-test coding task. It completed the tool loop and left the tests passing.

I have not independently verified every coding agent. I would rather publish one reproducible result than put a long compatibility matrix on the page and pretend every green check means the same thing.

When OpenRouter is still the better choice

If you need Claude, GPT, Gemini, Llama, and hundreds of provider routes behind one account, OpenRouter is the obvious fit. Vancine is for the narrower case: you mainly want current Chinese frontier models, a short catalog, and lower prices on the compared paid listings.

I am still deciding what a smaller provider needs to publish before developers will trust it with production traffic. Uptime history? Rate-limit details? Version pinning? Better cost exports?

If you work with these models, I would like to hear which one matters most.

DEV Community