How AI API token pricing works and how to estimate your bill

What is a token?

A token is a piece of text the model reads or writes: a short word, part of a longer word, a number or a punctuation mark. As a rough guide for English text, 1,000 tokens is about 750 words, or one token is about four characters. Code, other languages and unusual formatting often use more tokens per word. Each model family uses its own tokenizer, so the same text can count slightly differently on different models.

You do not need to count tokens yourself. Every API response includes a usage object with the exact numbers, and the console shows them for every call.

Input tokens vs output tokens

Input tokens: everything you send in the request, including the system prompt, the full conversation history, any documents or code you paste, and tool definitions.
Output tokens: everything the model writes back, including reasoning text for models that show it.

Output tokens are usually priced higher than input tokens, often several times higher. So a request that sends a lot and gets a short answer (classification, extraction) is cheap per call, while a request that asks for long text (articles, full files of code) is driven mostly by the output price.

The formula

cost of one call = (input_tokens x input_price + output_tokens x output_price) / 1,000,000

monthly cost = cost of one typical call x calls per day x days

Worked example with example numbers, not real prices: suppose a model costs 3 units per 1M input tokens and 15 units per 1M output tokens. A call with 2,000 input tokens and 500 output tokens costs (2,000 x 3 + 500 x 15) / 1,000,000 = 0.0135 units. At 1,000 calls a day for 30 days, that is about 405 units a month. Replace the example prices with the real ones from the table below.

Real prices now

Estimated from the prices on the models page: 2,000 input tokens and 500 output tokens per chat. The console shows the actual charge.

Models	Per chat	1,000 chats
GEM 3.8 Flash	Rp 4	Rp 4.434
TT-6 Luna	Rp 11	Rp 10.642
TT-6.1 Terra	Rp 7	Rp 7.263
HK-5.5	Rp 3	Rp 3.487
GEM 4 Argon	Rp 85	Rp 84.649
TT-6.1 Sol	Rp 15	Rp 14.526

The live prices are also on the models page. In the console, Usage lists your last 50 calls with model, input tokens, output tokens and the exact cost, so you can check your estimate against real calls.

How to estimate your own bill step by step

Pick a model from the models page and note its input and output price per 1M tokens.
Send 5 to 10 realistic requests and read usage.prompt_tokens and usage.completion_tokens from each response (or check Usage).
Average the input and output tokens to get one typical call.
Put those averages into the formula above to get the cost per call.
Multiply by your expected calls per day and days per month.
Add a margin for retries and longer conversations, then top up a little more than that so the balance does not run out mid-month.

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.macloudai.com/v1", api_key=os.environ["MACLOUD_API_KEY"])
r = client.chat.completions.create(
    model="gem-3.8-flash",
    messages=[{"role": "user", "content": "Summarize: ..."}],
    max_tokens=200,
)
print("input tokens:", r.usage.prompt_tokens)
print("output tokens:", r.usage.completion_tokens)

# fill in real prices per 1M tokens from the models page
INPUT_PRICE, OUTPUT_PRICE = 0, 0
cost = (r.usage.prompt_tokens * INPUT_PRICE + r.usage.completion_tokens * OUTPUT_PRICE) / 1_000_000
print("cost of this call:", cost)

Why long conversations get expensive

Chat APIs are stateless: the model does not remember earlier messages, so your app sends the whole conversation again on every turn. If each turn adds about 500 tokens, turn 1 sends 500 input tokens, turn 10 sends about 5,000, and turn 40 about 20,000. The total input over a long chat grows much faster than the number of messages.

Coding agents such as Claude Code and Cursor agent mode are an extreme case, because they also send file contents and command output. That is why a single long agent session can cost more than hundreds of short API calls.

Start a new conversation when the topic changes.
Keep only the last few turns, or replace older turns with a short summary.
Keep the system prompt short; it is resent on every call.
Send only the parts of a document that the question needs.

What does max_tokens do?

max_tokens is the maximum number of output tokens the model may generate for one reply. It is a ceiling, not a target: you pay for the tokens actually generated, not for the limit. Setting it protects you from runaway answers, for example a loop that keeps writing, or a model that writes three paragraphs when you wanted a single label. If a reply is cut off in the middle, the limit was too low; raise it for that kind of request.

Other ways to spend less

Use a smaller model for simple jobs and a large one only where quality clearly improves. See which model for which job.
Ask for short, structured output (JSON with fixed fields) instead of free text.
Cache answers in your own app for questions that repeat.
Create one key per project so the usage log shows which project spends what.

Billing is prepaid: you top up a balance, each call is deducted after it finishes, and there is no monthly fee or minimum spend. The balance does not expire while your account is active.

FAQ

Q1How many words is 1 million tokens?

For English text, roughly 750,000 words. Code and other languages usually need more tokens per word, and each model family counts slightly differently.

Q2Why are output tokens more expensive?

Generating text takes more compute than reading it, so providers price output higher. Short answers keep the output part of the bill small.

Q3Am I charged for max_tokens even if the answer is shorter?

No. You pay for the output tokens the model actually generates. max_tokens is only an upper limit.

Q4Why did my bill jump with long chats?

The whole conversation is resent on every turn, so input tokens grow with each message. Start new conversations or trim the history.

Q5Where can I see the cost of each call?

In the console under Usage: the last 50 calls with model, input tokens, output tokens and cost.