From Four to Nine: A Practical Gateway for Chinese AI Models

The useful announcement is not that another model directory became larger. The useful announcement is that an overseas developer can now keep one OpenAI-compatible integration while reaching nine Chinese AI provider families through one USD account and one operational boundary.

AIWave's live catalog returned 63 model records on 2026-08-27. The public positioning is therefore 60+ AI models from 9 Chinese AI providers: DeepSeek, GLM/Zhipu, Kimi/Moonshot, ERNIE/Baidu, MiniMAX, Qwen/Alibaba, Doubao/ByteDance, StepFun, and Xiaomi MiMo.

That distinction matters for production teams. A larger menu is not automatically a better system. The value is the ability to compare a small acceptance set, switch the model parameter when the task demands it, retain the same client contract, and keep a dated usage ledger in USD. The gateway is for teams that already know what they are building and need a dependable path to Chinese model routes. It is not positioned as an everything-for-everyone platform.

What changed: four provider families became nine

The first generation of the catalog concentrated on DeepSeek, GLM, Kimi, and ERNIE. The current catalog adds five provider families with distinct engineering tradeoffs:

The other four families remain important because they give a team a stable baseline: DeepSeek for a familiar reasoning comparison, GLM and ERNIE for Chinese ecosystem coverage, and Kimi for long-context and tool-use evaluations. The point of nine providers is not to claim that one route wins every benchmark. The point is to make the comparison operationally cheap in engineering time: one client contract, one account boundary, one audit trail.

Price table: separate verified dollars from live gateway ratios

Price claims age quickly. On 2026-08-27, the public AIWave endpoint returned 63 records with no duplicate model names. Its model_ratio, completion_ratio, and cache_ratio fields are gateway billing ratios, not a universal USD formula. The reconciliation script therefore reports those fields without converting them into dollars.

The following table contains dated, directly stated reference prices. The StepFun rows are the AIWave USD rate-card rows checked on 2026-08-27. The MiniMAX, Qwen, and Doubao rows are upstream provider reference prices and are labeled as such; they are not silently presented as AIWave invoice rates.

Sources: MiniMax token plan, BytePlus ModelArk pricing, Alibaba Cloud Model Studio pricing, and StepFun pricing. The AIWave source of truth for live model IDs is /api/pricing.

Two rules follow from the table. First, do not turn a gateway ratio into a dollar number by intuition. Second, do not mix an upstream provider price with an AIWave invoice rate. A procurement sheet should have separate columns for source, currency, rate_date, model_id, uncached_input, cache_hit, and output.

A concrete cost ledger

Assume a small agent workload produces 10M uncached input tokens, 2M cache-hit input tokens, and 1M output tokens in a month. Using the dated AIWave 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 difference is $3.77 for this particular token mix. That is not a model quality verdict. It is a reason to measure the acceptance set, output length, retry rate, and cache share instead of selecting by a single headline number.

A ledger row should look like this:

Keep retries and fallback calls as their own rows. Otherwise a model switch can appear to be a rate change when it is actually a reliability or prompt-shape change. For a paid engineering team, that auditability is more useful than a larger list of untested model names.

Runnable OpenAI-compatible example

The client contract stays familiar. Replace the model ID after testing it against your own prompts, not by changing SDK code:

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 invoice number and total from this text."},
    ],
    temperature=0,
)

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

Before production, pin the exact ID, set an output cap, record response.usage, and compare a fixed evaluation set across two or three routes. Do not put a real key in source control. The API page is the live catalog check; the dated pricing page is the rate reference.

Choosing the first five tests

Start with workflows that produce a verifiable result:

This test list is deliberately narrow. It creates reusable evidence for a team that pays for production capacity. It also remains searchable after a model launch announcement has stopped being new.

What the gateway is—and is not

AIWave is a unified USD gateway for overseas developers who need access to Chinese AI routes through a consistent OpenAI-compatible interface. The core promise is a single integration boundary and a clear rate-verification habit.

It is not a promise that every model is suitable for every task. It is not a claim that a single benchmark settles procurement. It is not a reason to hide cache pricing, silently translate currencies, or publish an exact count without a live check. The catalog can change, so pin model IDs and keep the date beside every price used in a forecast.

For the current list, check /models/. For the current API metadata, check /api/pricing. For the trust and operational boundary, check /trust. The durable value is not launch-day attention; it is the trust asset created when a developer can reproduce the route, the cost, and the result three months later.

Five-question publication check

Before this article goes live, the answer is yes to all five questions: