Cheap LLM API Provider: What to Check Beyond Price
A cheap llm api provider is only cheap if the whole bill is low: output rates, failed calls, context size and expiring credit all count. Use this checklist, with worked costs at $0.15 per 1M input and $0.50 per 1M output tokens.
Updated

Cheap AI APIs
Key points
- Compare input and output prices separately and use your own token ratio.
- Check the context window, rate limits, billed errors, expiry and data use.
- Summaries cost $3.25 per 1,000 documents here; 1,000 generated stories cost $1.575.
- A $0.50 trial needs no card and lasts 7 days.
Price per million tokens is only half the quote
Picking a provider by the lowest number on a pricing page works until the first real bill. Two providers can advertise the same input price and still cost you very different amounts, because of how output is priced, what counts as a billable request, and how much you spend retrying.
This article gives you a short checklist you can apply to any provider, then runs one as an example: $0.15 per 1M input tokens and $0.50 per 1M output tokens, the rates on this site. Use the checklist on others too. For their numbers, read their current pricing page, because prices change.
The checklist
| Check | Why it matters |
|---|---|
| Input and output prices separately | A blended price hides which side you pay most on |
| Context window | A small window forces chunking, which means more calls and repeated instructions |
| Are failed requests billed? | Errors and refusals that cost money inflate the bill |
| Rate limits | Low concurrency turns a cheap batch into a slow one |
| Minimum top-up and expiry | Credit that expires is money you lose |
| Subscription or prepaid | A fixed monthly fee punishes light months |
| Which model is it, honestly? | Cheap can mean a different or smaller model than the name suggests |
| Data use | Are prompts used for training? |
| API compatibility | OpenAI-compatible means no rewrite and easy switching |
Write the answers for each provider in a small table of your own. Gaps in a provider's documentation are an answer too.
Input versus output: the ratio decides
Output tokens usually cost more than input tokens, so what your app does matters. Here are two workloads at $0.15 per 1M input and $0.50 per 1M output.
Summaries. A document of 20,000 tokens goes in and 500 tokens come out. Input: 20,000 x $0.15 / 1,000,000 = $0.003. Output: 500 x $0.50 / 1,000,000 = $0.00025. Total $0.00325 per document, or $3.25 for 1,000 documents. The input side is 92% of the bill.
Generation. A 500-token prompt produces 3,000 tokens of story. Input: $0.000075. Output: $0.0015. Total $0.001575, or $1.575 for 1,000 generations. Here output is 95% of the bill.
A provider with cheap input and expensive output wins the first workload and loses the second. Compare on your ratio, not on a headline figure. For a mixed month of 10M input tokens and 2M output tokens, the bill here is $1.50 + $1.00 = $2.50.
Errors, retries and what you actually pay
The cheapest provider on paper becomes expensive if a share of your calls fail and you pay for them. Ask whether failed or refused requests are billed. Here they are not. Ask how rate limiting behaves: a full key allows 8 concurrent requests and 300 requests per minute, a trial key 2 and 60. Over a limit the API returns HTTP 429 with the code rate_limited or concurrency, and on temporary overload also 429 with Retry-After: 5.
Handle that politely and the cost of retries is zero, since the rejected request never ran. Here is a minimal wrapper:
import time
from openai import OpenAI, RateLimitError
client = OpenAI(base_url="https://api.cheapaiapis.com/v1", api_key="YOUR_KEY")
def ask(prompt, retries=4):
for i in range(retries):
try:
r = client.chat.completions.create(
model="uncensored",
messages=[{"role": "user", "content": prompt}],
)
return r.choices[0].message.content
except RateLimitError:
time.sleep(5)
raise RuntimeError("rate limited after retries")
Streams occupy their concurrency slot until they finish, so a batch job that streams 20 responses at once will be limited by the 8 slot cap, not by the per minute figure.
Context window and the hidden cost of chunking
A 100,000-token window, shared by prompt and completion, lets you send a long document in one request. With a small window you split it, repeat the instructions in each chunk, and add a final call to merge the results. That is more requests, more input tokens and more code to maintain. If your workload involves long inputs, check the window before the price.
The same applies to chat apps with long histories. If history fits, you can keep it simple. Even so, resending history costs input tokens each turn, so trimming is still worthwhile. A related read is cheap GPT API myths versus facts, which separates real savings from marketing.
A month of real usage, priced
Put the pieces together with a small product: a writing assistant with 300 active users who each send 40 requests a month. That is 12,000 requests. Each sends about 3,000 input tokens (instructions, story so far) and returns 700 tokens.
- Input: 12,000 x 3,000 = 36M tokens, which costs 36 x $0.15 = $5.40.
- Output: 12,000 x 700 = 8.4M tokens, which costs 8.4 x $0.50 = $4.20.
- Total: $9.60 a month, or $0.032 per user.
Now see what small changes do. Trimming history so input drops to 2,000 tokens per request saves 12M input tokens, or $1.80. Capping output at 500 tokens saves 2.4M tokens, or $1.20. A provider that bills failed calls, with a 5% failure rate, adds 5% to the whole bill. Those effects are as large as a price difference between two cheap providers, which is why the checklist matters.
The model is the other half of the comparison. A lower price for a model that needs three tries to follow your format is not lower. Before you decide, run 20 of your real prompts and count how many give a usable answer the first time.
Prepaid, expiry and top-up rules
Look for two policies: whether credit expires, and what the minimum payment is. Here credit never expires, there is no subscription, and top-ups are any whole amount from $20 to $500. The bonus is +5% from $50, +10% from $100 and +15% from $200, so a $100 top-up becomes $110 of credit. You can pay with crypto (USDT on TRC20, USDC on Base). The pricing page has the table.
A $20 top-up at the generation workload above ($0.001575 per story) buys about 12,698 stories. At the summary workload ($0.00325 per document) it buys about 6,153 documents. These are the figures to compare against any other provider's equivalent.
Test it before you commit
The best check is your own prompts. Sign up with Google, or with email and password, and claim the $0.50 trial credit on the API key page. It takes one click, needs no card, lasts 7 days and is limited to one per person. Since the API is OpenAI-compatible, the test is a base URL change.
curl https://api.cheapaiapis.com/v1/chat/completions \
-H "Authorization: Bearer YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"uncensored","messages":[{"role":"user","content":"Summarize: retries need backoff."}]}'
Every response carries X-Balance-USD, so you can watch the real cost of your test run. Be clear about what is on offer: one open-weight model, run by an independent provider and tuned to answer without content refusals for lawful adult use. It is not the official API of any model vendor, and it is text only, with no images, embeddings or fine-tuning. Prompts are not used for training. Details are in the docs.
More guides
Questions and answers
Read the docsIs the lowest price per million tokens always cheapest?
No. Output pricing, billed failures, small context windows and expiring credit can raise the real bill. Compare on your own input and output ratio.
Are failed requests charged here?
No. Failed or refused requests are not charged, and credit never expires.
How do I test a provider cheaply?
Use the trial credit or a small top-up and run your real prompts. Here a trial of $0.50 needs no card and lasts 7 days.
Is this the official API of a big model vendor?
No. It is one open-weight model run by an independent provider, not the official API of any vendor and not a router.
Your key is one form away
Create an account, copy the key, change the base URL. That is the whole setup.
https://api.cheapaiapis.com/v1