1) A request = ONE call your program makes to the AI

A request is a program asking the AI once and getting one answer back. CAREFUL: one message you type is NOT necessarily one request. How many requests your message becomes depends on the tool you use.

Program asksModel thinksModel answers

= 1 REQUEST

Whether your message is 5 words or 500, that single call counts as 1 request. Length does not change the NUMBER of requests — but it does change the tokens burned, and therefore what you pay under PAYG.

So how many requests does one message of yours become?

In a chat window: 1 message = 1 request. Simple. But in coding tools (Claude Code, Cline, Codex…) it is very different. These tools work like agents: you write one sentence, and they work step by step in the background — asking the model again at EVERY step.

Example: you told Claude Code “fix this bug”. Behind the scenes:

1It reads your files → 1 request

2It studies the code to find the bug → 1 request

3It writes the fix → 1 request

4It runs the tests and reads the result → 1 request

5If a test failed, it fixes again → 1 more request… and so on

So your ONE sentence can become 20–30 requests in the background. That is exactly why a “1,000 requests/day” package runs out faster than you expected. This is not a bug — it is how agentic tools work.

You can always see how many requests you used in the Activity page of your panel; every request is its own line.

2) Tokens = pieces of words

The AI does not read words the way we do. It splits text into small pieces, and each piece is called a TOKEN. A token is sometimes a whole word, sometimes half a word, sometimes a punctuation mark.

This sentence: “The weather is lovely today.”

Bugün hava çok güzel.

That is about 5 tokens. Roughly: 1 token ≈ 4 characters.

Pricing is based on tokens — not on words or characters.

3) Two kinds of tokens: going out and coming back

Tokens flow in both directions in a request. Both are counted.

INPUT (what you send) — your question + the instructions given to the AI + any earlier conversation.
OUTPUT (what the model writes) — the whole answer you get back. Output is usually more expensive than input, because that is the part that really costs the model work.

⚠ The most surprising part: in a chat, the history is re-sent every single time

The AI has NO memory. It does not remember your previous message. So when you continue a chat, the program re-sends the ENTIRE conversation from the start — otherwise the model would not know what you are talking about. Which means: the longer the chat, the more expensive each message gets:

1Message 1 — only your question is sent~30 token

2Message 2 — question 1 + answer 1 + question 2 are sent~90 token

3Message 3 — all of the above + question 3 are sent~200 token

What to do: start a NEW CHAT when the topic changes. Letting 50 messages pile up in one chat makes every new message very expensive.

4) In coding tools, even “hello” never travels alone

Tools like Claude Code and Cline send a lot in the background without you noticing: the contents of your open files, the project structure, the instructions given to the tool… So even typing “hello” can send thousands of tokens behind the scenes.

This is not a bug — it is how the tool works, and it explains why even a small question costs something.

5) How is it counted? Packages and balance work DIFFERENTLY

📦 If you bought a package

You get a daily allowance. Each REQUEST (each API call) uses 1 credit — not each message you type. Short or long makes no difference. With a 500 requests/day package, an agentic tool can burn that in a handful of tasks.

1 API call = 1 credit

💳 If you pay from balance (PAYG)

No credits. Whatever tokens a request burns is what comes off your balance. Short question = little money. Long chat = a lot of money.

pay per token

Real example: on claude-haiku-4-5, 1M tokens cost $0.3. So a 1,000-token request costs about 0.0142 ₺. (Prices from the live catalog.)

6) Why do I see THREE lines in the ledger?

In Activity you will see three lines for a single request. Do not panic — the money leaves once. Here is how it works: at the start we cannot know what a request will cost, so we first “set aside” an amount. When it finishes we give back ALL of what we set aside, and deduct only what was actually spent.

Reservation — set aside (estimated maximum)−$0,05

Refund — the whole reservation returned+$0,05

Usage — what was actually spent, deducted−$0,01

Net result: only what was actually spent leaves your pocket. All three carry the same request ID, so nothing gets mixed up.

7) How do I spend less?

Start a new chat when the topic changes — so the old conversation is not re-sent every time.
Be specific. Say exactly what you want instead of explaining at length.
Use a cheap model for simple work (GLM, Kimi). Not every job needs the most expensive model.
If you use it regularly, buy a package — being credit-based, it works out far cheaper on long requests.

YARDIM MERKEZİ

Sık sorulan sorular

Ücretlendirme, paketler, anahtarlar ve kapalı yüzeyler hakkında kısa ve dürüst cevaplar. Emin olmadığımız hiçbir şeyi kesinmiş gibi yazmıyoruz.

12 / 12 soru gösteriliyor. Doğrulanmamış hiçbir cevabı kesinmiş gibi yazmıyoruz — onay bekleyenler açıkça etiketlidir.

Cevabını bulamadın mı?

Destek kanalı adresleri doğrulanmadan yayınlanmaz — sahte link göstermiyoruz. Şimdilik en hızlı yol: dokümandaki hata kodları ve yanıttaki istek kimliği.