EN ▾

Chinese LLM APITranslation

Get API key

Updated

Cheap Chinese LLM API: What to Compare Beyond Price

Per-token price is only half of what you pay. Measure tokens on your own text, then compare context, limits and minimums.

Abstract brush-stroke tiles linked to a price-tag crystal

Start with the cost of your own text

A cheap chinese llm api is only cheap if it handles your text at a fair price per request. Price lists show cost per million tokens, but a token is not a character or a word. How many tokens a Chinese sentence needs depends on the tokenizer of the model behind the API, and two providers can bill very different totals for the same paragraph even at identical per-token rates.

So the first step is to measure. Send a sample of your own Chinese prompts, read the usage field in the response, and compute the cost per request. Our rates are $0.15 per 1M input tokens and $0.50 per 1M output tokens, and the code below turns a sample into an average cost.

from openai import OpenAI

client = OpenAI(base_url="https://api.chinesellmapi.com/v1", api_key="YOUR_KEY")
samples = ["请用三句话介绍一下杭州。", "把下面这段话翻译成英文:今天天气很好。"]

for text in samples:
    r = client.chat.completions.create(
        model="uncensored",
        messages=[{"role": "user", "content": text}],
    )
    u = r.usage
    cost = (u.prompt_tokens * 0.15 + u.completion_tokens * 0.50) / 1_000_000
    print(len(text), "chars ->", u.prompt_tokens, "in,", u.completion_tokens, "out, $%.6f" % cost)

Run 30 to 50 representative prompts rather than two. The ratio of tokens to characters you see is the number to use when you compare providers.

Compare input, output and minimums

The other cost factors are the same as for any English-language API, and they matter just as much with Chinese text.

  • Input and output prices. Output usually costs more per token. Here it is $0.50 against $0.15 for input, so a task that generates long replies, such as story writing, costs more than one that classifies short texts.
  • Minimum top-up. Ours starts at $20 and goes up to $500 in whole amounts.
  • Expiry and subscriptions. Credit here never expires and there is no subscription. Check both on every provider you compare.
  • Failed requests. Failed or refused requests are not charged here. Not every provider does this.

For other providers, check their current pricing instead of relying on a blog post, including this one. Prices change.

Language quality matters more than a small price gap

If your users read Chinese, quality problems show up as mixed-language answers, stiff phrasing, wrong measure words and slips in tone. A model that is a few cents cheaper per million tokens but needs two attempts per answer is more expensive in practice.

Build a small evaluation set in the style your product needs:

  1. Ten prompts in everyday spoken Chinese, ten in formal writing, ten in the genre you serve, such as customer support or web fiction.
  2. Ask a native reader to rate each answer for naturalness and accuracy on a 1 to 5 scale.
  3. Check that the model answers in Chinese when asked in Chinese, and keeps to Simplified or Traditional characters as instructed.

Put the script and style in a system prompt, for example a note that tells the model to answer in Simplified Chinese, in a natural tone. The service runs one open-weight model, so you cannot switch models to repair quality. You adjust the prompt, and if the results are not good enough, you choose another provider. That is a fair trade-off to know about before you start.

One more check is worth the ten minutes: ask the same question three times and compare. Sampling makes answers vary, and for a product you want to know how wide the spread is. Setting a seed and a low temperature narrows it, which helps when you tune a prompt and want to see only your own changes.

Context length and long Chinese documents

Context is counted in tokens, not characters. The window here is 100,000 tokens, prompt plus completion, and a completion can use whatever remains. For long documents such as contracts, novels or chat logs, check how many tokens your sample takes with a real call before you count on the limit. The same page of Chinese text can use fewer or more tokens than its English version, depending on the tokenizer.

Long inputs also cost money every time you send them. A 60,000 token document costs 60,000 x $0.15 / 1M = $0.009 per read. If you ask ten questions about it, send it ten times, and you pay $0.09 in input alone. For repeated questions, split the work: summarize once and ask questions against the summary.

Limits that affect a cheap setup

A low price does not help if rate limits stall your app. A full key allows 8 concurrent requests and 300 requests per minute. A trial key allows 2 and 60. Over a limit the API returns HTTP 429 with the error code rate_limited or concurrency. On temporary overload it returns 429 with a Retry-After: 5 header.

A streamed reply holds its concurrency slot until the stream ends, which is easy to forget in a chat app where users read slowly. Plan for a queue in front of the API if you expect more than 8 people to chat at the same moment. The maximum request body is 8 MB, which is far above what a text prompt needs.

Common Chinese-language use cases and their cost profile

Different jobs sit at different points on the price scale, so match the test to the job.

  • Customer support chat. Short questions, medium answers, a long system prompt. Input dominates, so keep the system prompt tight.
  • Translation. Input and output are about the same length. Our Chinese-English translation API page shows the pattern.
  • Story and roleplay generation. Short prompts, long replies. Output dominates, so the $0.50 rate is the one to watch.
  • Classification and tagging. Very short outputs. These are the cheapest jobs per request, and a good fit for JSON mode.

For a first working app, the Chinese language app quickstart has step-by-step code. Set temperature lower for translation and tagging, and higher for fiction. Most clients pass those settings through unchanged.

What this service is, and what it is not

It is a hosted, OpenAI-compatible text API running one open-weight model from an independent provider. The model is tuned to answer without content refusals for lawful adult use. That suits fiction, roleplay and companion apps for adults, including Chinese web fiction and roleplay. Sexual content involving minors is never allowed.

It is not the official API of any model vendor, and it is not a router. There is no embeddings endpoint, no image, audio or video generation and no fine-tuning. If you need embeddings for Chinese search, pair this API with another service for that part.

Prompts are not used for training. Streaming, tool calling and JSON mode (response_format with json_object) all work. Any OpenAI SDK connects when you set the base URL: change the base URL to https://api.chinesellmapi.com/v1, keep your key, and set the model to uncensored.

How to try it at no cost

Sign up with Google or email, open the API key page and claim $0.50 of trial credit with one click. No card is needed, it is one claim per person, and it is valid for 7 days. At 2,000 input and 500 output tokens per request, which costs $0.00055, the trial covers about 909 requests.

After that, top-ups are any whole amount from $20 to $500, by crypto (USDT on TRC20, USDC on Base). Bonuses are +5% from $50, +10% from $100 and +15% from $200. The pricing page has the full table, and the docs list every parameter.

Questions and answers

How many tokens does Chinese text use?

It depends on the model's tokenizer. Send a sample and read the usage field to get your own ratio.

Can I test Chinese quality before paying?

Yes. Claim $0.50 of trial credit with one click on the API key page, with no card, valid for 7 days.

Is this the official API of a Chinese model vendor?

No. It runs one open-weight model from an independent provider and is not a vendor's official API.

Does it offer embeddings for Chinese search?

No. It is a text chat API only, so use another service for embeddings.

Your key is one form away

Create an account, copy the key, change the base URL. That is the whole setup.

Get API key