What that request costs
Rate card checked 2026-10-24. DeepSeek V4 Pro: $1.32 per million input tokens, $3.96 output. V4 Flash: $0.30 and $1.20. Cached input on Pro: $0.044.
Minimum top-up is $5, billed in USD. Checkout currently lists PayPal.
Chinese AI API: One Endpoint for DeepSeek, Kimi, GLM and Qwen
One key, one USD bill, one published rate per model, and no client rewrite. DeepSeek V4 Pro costs $1.32 per million input tokens here. That is the rate we charge today, not a launch price.
The whole integration is one line.
You don't want five accounts, five SDKs and five invoices to call four Chinese model families. So don't build it: point the OpenAI client you already have at one base URL, pass any model ID, and read one bill in USD.
OpenAI-compatible chat completionsCopy-ready request
curl https://aiwave.live/v1/chat/completions \
-H "Authorization: Bearer $AIWAVE_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"deepseek-v4-pro","messages":[{"role":"user","content":"Explain this stack trace."}]}'What is a Chinese AI API?
A Chinese AI API is an HTTP endpoint that serves models built by Chinese labs (DeepSeek, Kimi, GLM, Qwen) in the same request format OpenAI uses. You send a chat completion to one base URL with a model ID and get JSON back. No local install, no separate SDK.
One base URL · many model IDsOpenAI compatibility is the easy part
Almost every Chinese lab now ships its own OpenAI-compatible endpoint, so the request format isn't where you get hurt. You get hurt on the other four things, each multiplied by the number of providers: signup and identity checks, balance top-ups, invoices, and rate limits. Add a status page per provider and you've built an on-call rotation for vendors you don't control.
Call it a Chinese AI API gateway, a router, or a relay. Same job: one door, many models, one place to see what you spent.
What one endpoint means here, and what it doesn't
It means one URL, one key, one bill, and a record of what each request cost.
It does not mean every model behaves identically. Parameter support varies: some models ignore temperature, tool-calling and JSON mode differ in small ways, and reasoning models return an extra reasoning_content field next to the answer. If a parameter has to behave the same way everywhere, test it per model before you ship. We'd rather tell you that now than have you find out from a 500 at 2am.
What it costs
Every row below is from the public rate card, generated from the live pricing endpoint. The card was checked 2026-10-24, and the per-row date is when that rate took effect. USD per million tokens, default plan: the rate we charge, not a percentage off somebody else's rate that moves on Monday.
USD per 1M tokens · default group ×1| Model ID | Input / 1M | Output / 1M | Cache hit / 1M | Rate date | Use it for |
|---|---|---|---|---|---|
| deepseek-v4-pro | $1.32 | $3.96 | $0.044 | 2026-10-20 | Code, debugging, multi-step reasoning |
| deepseek-v4-flash | $0.30 | $1.20 | $0.006 | 2026-10-20 | Extraction, tagging, summaries, volume |
| glm-5 | $1.55 | $4.96 | $0.40 | 2026-08-27 | Structured output, Chinese-English, lower rate |
| glm-5.1 | $2.10 | $6.60 | $0.68 | 2026-08-27 | Structured output, Chinese-English, higher rate |
| kimi-k3 | $4.50 | $22.50 | $0.90 | 2026-08-27 | Long documents, repo-sized context |
| qwen3.8-max | $2.678 | $8.034 | Not listed | 2026-08-27 | Multilingual generalist, second opinion |
Source: aiwave.live/pricing, dated rate card generated from /api/v1/pricing. Qwen's cache-hit rate isn't on the card, so the cell says so rather than carrying a guess.
One flat price, all week.
The number you see at 2pm is the number you see at 2am. We don't discount off-peak and we don't add a peak premium, because a bill that moves with the clock is its own kind of cost. What you're buying is predictability, not a number that resets at midnight. The next block says exactly when that trade is a bad one for you.
The number you see at 2pm is the number you see at 2am. We don't discount off-peak and we don't add a peak premium, because a bill that moves with the clock is its own kind of cost. What you're buying is predictability, not a number that resets at midnight. The next block says exactly when that trade is a bad one for you.
If you came here looking for a cheap Chinese AI API
Here is the answer we'll sign: V4 Flash at $0.30 per million input tokens is genuinely cheap, Kimi at $4.50 is not, and anyone quoting a large percentage off is comparing one hour of their day to a different hour of yours.
The part most price tables leave out
DeepSeek's own price list doesn't work like ours. It has a peak window, 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, and everything outside it bills at half the peak rate. That window is 35 hours out of a 168-hour week, and weekends are off-peak all day. Our DeepSeek rates equal DeepSeek's peak rate; on the VIP and team plans they're 10% below DeepSeek's peak-hour rate.
Which means, stated plainly: during peak hours you pay the same, or a little less, and for 79% of the week (off-peak hours plus all weekend) going direct with a card DeepSeek accepts costs about half what we charge. If your workload is batch jobs you can schedule into that window, go direct. There's a section below that says it again.
A day of real work, priced
1,000 requests a day at 2,000 input and 800 output tokens each on V4 Pro is 2M input and 0.8M output: $2.64 plus $3.17, or $5.81 a day.
Now assume 70% of input is a cache hit, which happens once your system prompt and retrieved context stop changing: $4.02 a day. Cache hits move this bill more than any plan decision. Cached input on V4 Pro is $0.044, the line item nobody reads until the invoice is wrong.
What actually moves your bill
Three things, in order: how much input you re-send, how long the output runs, how often you retry. Output costs 3× input on V4 Pro, so "a full explanation" is three times the price of an answer with a reason. No plan or router fixes a prompt that asks for 800 tokens when 200 would do.
Plans
Default is the rate card above. VIP, svip, coding team and Visual Team plans bill at 0.9× those rates. If you're running a coding agent all day, that's the one to ask about.
Start with $5. No subscription, no monthly commitment. Confirm the method and the amount at checkout before you pay.
Which models can you call
Each model below gets the same four answers: what it's good at, what it isn't, what it costs, and the exact string to pass. Read them as routing advice, not benchmarks. The model pages record what has actually been verified, and for most routes that is still nothing.
Six routes · dated rates · exact IDsDeepSeek V4 Pro: the default for code
Start here if you're writing code, reading a stack trace, or asking for a plan you'll actually follow. It holds a spec across a long conversation better than Flash, which is why we'd put it behind a coding agent. It's the wrong tool for high-volume classification: you'd pay 4.4× the Flash rate for a task that doesn't need reasoning.
$1.32 per million input, $3.96 output, $0.044 cached. "model": "deepseek-v4-pro"
DeepSeek V4 Flash: the workhorse
Flash is for tasks with a right answer that you need ten thousand times: extraction, tagging, short summaries, routing decisions. It's cheap enough that batching stops being worth the engineering. What it won't do is hold a long chain of reasoning without drifting, so don't hand it a 40-step agent loop and then blame the router.
$0.30 in, $1.20 out, $0.006 cached, about a quarter of Pro on input. "model": "deepseek-v4-flash"
GLM-5 and GLM-5.1: structured output and Chinese-English
GLM-5.1 is the one to test when the output has to parse on the first try: JSON schemas, function calls, form filling, translation that keeps its formatting. We have not verified those capabilities on this route, so send your own schema through it before you build on it. If it holds, $2.10 in is cheap next to the retry logic it saves. glm-5 is the same family at $1.55 and $4.96 if the budget is tighter.
"model": "glm-5.1" or "model": "glm-5"
Kimi K3: long documents
Kimi is for the job where the input is the problem: a 300-page contract, a repository, a support archive you want to query in place. Test it when nothing else fits. It's also the most expensive per token of the six, so it's a poor default for chat. $4.50 in and $22.50 out: run the arithmetic before you point a chatbot at it.
"model": "kimi-k3"
Qwen3.8-Max: breadth and batch
Qwen is the generalist: multilingual, solid at code, comfortable with long inputs, priced at $2.678 in and $8.034 out, above Pro and well below Kimi. It's a useful second opinion when DeepSeek gives you an answer you don't trust. Where it loses is consistency on narrow, high-precision tasks: it will occasionally answer a slightly different question than the one you asked.
"model": "qwen3.8-max"
Also on the same key
GPT, Claude and Gemini routes are listed on the pricing page as Verification pending: visible in the catalogue, not yet priced. We publish a rate only after it reconciles against the provider's own price sheet, so there is no row to quote here yet. When those rows settle, they will appear with a date.
Migrate in one line
Two changes: base URL, and model ID. That's it.
Python · Node · any client with a base URLPython
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["AIWAVE_API_KEY"],
base_url="https://aiwave.live/v1",
)
response = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[{"role": "user", "content": "Summarize this error log."}],
)
print(response.choices[0].message.content)Node
import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.AIWAVE_API_KEY,
baseURL: "https://aiwave.live/v1",
});
const response = await client.chat.completions.create({
model: "deepseek-v4-pro",
messages: [{ role: "user", content: "Summarize this error log." }],
});
console.log(response.choices[0].message.content);Anything that takes a base URL
LangChain, LlamaIndex, the Vercel AI SDK and everything else that accepts a baseURL work the same way. If your client can set a base URL, it works.
Before you move traffic
What breaks, and how to catch it in one request.
Three failures account for most of the support tickets we see after a migration.
- Reasoning models return two fields. On
deepseek-v4-proanddeepseek-v4-flashyou may getreasoning_contentalongside the answer. Readchoices[0].message.content; if you're rendering the raw object, your users will see the model thinking. - Model IDs are exact strings.
deepseek-v4-prois notdeepseek-pro. Copy the ID from the table above rather than typing it from memory. - Tool-calling and JSON mode differ per model. Test the model you're switching to with your actual schema before you move traffic. This is the failure that shows up in week two, not hour one.
What a direct account doesn't give you
A direct account with each lab is not a worse product. It is a different set of chores, and these are the four that stop being chores.
Route, account, response, recordOne invoice
Every lab bills in its own currency on its own cycle. You get one USD statement, with the charge for each request attached to it.
One key to rotate
Key hygiene multiplies by however many providers you sign with, and it's the work that gets skipped first.
A model swap that costs nothing
Changing deepseek-v4-pro to qwen3.8-max is a string edit, not a new signup and a new billing relationship.
A place to check before you commit
Rate before the request, cost after it, without opening a support ticket to find out what happened.
And two things it doesn't give you either.
It does not give you lower latency than the provider's own region, and it doesn't invent features upstream doesn't have.
It does not give you lower latency than the provider's own region, and it doesn't invent features upstream doesn't have.
When AIWave is the wrong choice
A page that only lists reasons to buy is one you shouldn't trust. The first three are hard limits; the last two are about where you are.
Five cases · stated before you sign upYour workload is schedulable batch
If you can run it at 03:00 UTC on a Tuesday, DeepSeek's off-peak rate is about half of ours. Set a cron job against the provider directly and keep the money.
You need a compliance certificate we don't hold
If your buyer requires a specific audit report, a data-residency guarantee, or a signed DPA with named subprocessors, go to a provider that has the paperwork. We won't pretend to certifications we don't have.
You need day-zero access to a new model
Aggregators trail upstream launches by hours to days while routes are tested. If being first matters, be first somewhere else.
You're calling exactly one model, forever
Then one account with that one provider is simpler than a router, and you don't need us in the middle.
Your volume is big enough to sign a contract
At enterprise scale, a direct agreement with a lab will beat per-token list pricing. We can't out-negotiate that, and we won't pretend a router is a procurement strategy.
If none of those five describe you, the rest of this page is the argument.
Questions developers ask before the first request
Six answers, each one checkable against the pricing page and the docs.
Answers match the pricing pageWhat do I need to open an account?
An email address or a GitHub account, and a $5 minimum top-up in USD. Checkout currently lists PayPal, and that screen is the authority on the method and the amount.
Can I use this from the US or EU?
Yes. It's a public HTTPS endpoint: no VPN, no proxy, no regional client. What we can't do is give you a data-residency guarantee or a compliance certificate we don't hold, so ask before you send production traffic rather than after.
How does this compare with OpenRouter's fees?
Different fee structure, not a lower price. OpenRouter's published FAQ says it charges 5.5% when you buy credits, with a $0.80 minimum, so a $5 top-up costs $5.80 there before you send a single request. We charge no fee on top-ups.
On inference, OpenRouter passes through the provider's rate, including DeepSeek's off-peak discount, which we don't match. Heavy off-peak batch on DeepSeek: they win. Small top-ups at peak hours: their $0.80 minimum is 16% of a $5 top-up, and at peak the per-token rate is the same either way, so the fee is the whole difference.
Can I keep my existing OpenAI SDK and code?
Yes. Change base_url to https://aiwave.live/v1 and the model string, then keep everything else. If your code sets a base URL anywhere, it will work.
Which model should I start with?
Start with deepseek-v4-flash to see the shape of the response cheaply, then move to deepseek-v4-pro once the prompt is stable. Ten imperfect calls on Flash teach you more than one careful call on Pro.
What happens when a model is retired or renamed?
You get notice on the affected ID and a dated migration note. Pin exact model IDs rather than aliases so a rename never reaches production silently. If a route is pulled with no notice, we tell you which requests failed.
Start with one request
Run one call you can inspect afterward: the response, the tokens, the cost. No card is needed to start.