From Four to Nine: One OpenAI-Compatible Route to Chinese AI Models

AIWave began with four Chinese provider families: DeepSeek, GLM/Zhipu, Kimi/Moonshot and ERNIE/Baidu. The live catalog now has 63 model records across nine families. MiniMAX, Qwen/Alibaba, Doubao/ByteDance, StepFun and Xiaomi MiMo are the five additions.

It is tempting to turn that into a logo announcement. A developer evaluating a gateway needs something less decorative: exact model IDs, a dated price boundary, a cost example and code that does not change when the provider does.

That is the useful part of moving from four to nine.

What changed in the catalog

The new families are not single interchangeable models.

• MiniMAX includes M2, M2.1, M2.5, M2.7, M3 and high-speed variants in the current catalog.
• Qwen spans Qwen3.8, 3.7, 3.6, 3.5, Max, audio, image and embedding routes. “Use Qwen” is not a reproducible instruction; pin the ID.
• Doubao routes are versioned, including Seed 2.0 and 2.1 variants. Keep the version in your test record.
• StepFun currently has step-3.5-flash, step-3.5-flash-2603 and step-3.7-flash.
• Xiaomi MiMo is represented by xiaomi/mimo-v2.5-pro in the live catalog.

The catalog can move after this article is published. Treat /api/pricing as the model-ID source of truth and keep the date beside any forecast.

Keep three kinds of price data separate

Gateway price discussions get unreliable when three different things are collapsed into one table:

1. an upstream provider's public rate;
2. a gateway's internal billing ratio;
3. the USD rate a customer can actually reconcile against an invoice.

They are related, but they are not the same number.

The AIWave endpoint exposes model_ratio, completion_ratio and cache_ratio for the live catalog. Those fields are useful for reconciliation, but they are not permission to invent a USD price. A USD row belongs in a cost forecast only when the dated AIWave rate card states it directly.

Here are the StepFun rows checked on 2026-08-27, per one million tokens:

Model	Uncached input	Cache hit	Output
step-3.5-flash	$0.21	$0.04	$0.63
step-3.5-flash-2603	$0.21	$0.04	$0.63
step-3.7-flash	$0.40	$0.08	$2.42

Upstream reference pages for MiniMAX, Doubao through BytePlus, Qwen and StepFun are still useful. Label their currency, region, context tier and source. Do not quietly present them as an AIWave invoice rate.

Work one cost example before you benchmark

Suppose a monthly agent workload produces:

• 10 million uncached input tokens;
• 2 million cache-hit input tokens;
• 1 million output tokens.

Using the dated StepFun rows:

step-3.5-flash = 10 × 0.21 + 2 × 0.04 + 1 × 0.63 = $2.81
step-3.7-flash = 10 × 0.40 + 2 × 0.08 + 1 × 2.42 = $6.58

The $3.77 difference is not a quality verdict. It tells you how much the same token shape costs before retries and fallbacks. Your ledger should keep those extra calls as separate rows; otherwise a route failure can look like a mysterious price change.

A useful record contains the workspace, exact model ID, uncached input, cache-hit input, output, rate date, status and request ID. That small amount of bookkeeping makes a later model switch explainable.

The client contract stays familiar

The provider changes through the model parameter, not through a new application integration:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://aiwave.live/v1",
    api_key=os.environ["AIWAVE_API_KEY"],
)

response = client.chat.completions.create(
    model="step-3.5-flash-2603",
    messages=[
        {"role": "system", "content": "Return compact JSON only."},
        {"role": "user", "content": "Extract the order ID and total from this text."},
    ],
    temperature=0,
)

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

Keep the key in an environment variable, set an output cap and store the returned usage. Before production, run the same acceptance set against two or three routes.

Five tests are more useful than 63 casual prompts

Start with work that has a checkable answer:

1. Structured extraction: validate the response against a JSON schema.
2. Chinese-English transformation: score names, numbers and formatting, not only fluency.
3. Long-context retrieval: verify cited spans from a fixed document pack.
4. Agent sub-task: bound output and record retries or malformed tool arguments.
5. Fallback: make one staging route unavailable and confirm that changing the model parameter is enough.

This is deliberately narrower than a general benchmark. A paid engineering team needs evidence for its own workload, not a universal winner.

That distinction matters because the first outside AIWave customers did not wander through the full catalog. Their calls were overwhelmingly DeepSeek V4 Pro, with high input volume. Breadth made comparison possible; a dependable route for a specific workload was what they paid for.

What nine providers does—and does not—mean

AIWave is a dollar-denominated gateway for developers outside China who need one OpenAI-compatible boundary into Chinese AI models. The promise is simpler account, routing and billing work.

It does not mean every model is ready for every task. It does not make an upstream price an invoice rate. It does not make a model family name precise enough for deployment. And it does not replace an acceptance test.

Use the current catalog, dated pricing ledger, observed route evidence and integration examples together. The model count gets old. A reproducible route, cost and result remain useful.