Buying AI tokens: how a dollar balance works and what $5 gets you

When people search for how to buy AI tokens, they often imagine a package: "I bought a million tokens, now I spend them". A gateway works differently, and understanding that removes a lot of questions. You top up a dollar balance, and every request deducts money from it at the price of the chosen model. Below we look at exactly what is charged and calculate how many requests $5 covers on different models.

Prices in this article are as of October 18, 2026; the current list is GET https://api.zukko.pro/v1/models or the "Models" page in the dashboard.

What you actually buy

You do not buy tokens of a specific model in advance. You put dollars on the account, and spending then goes like this:

1. You send a request to a model, for example claude-sonnet-5-5.
2. The model answers, and the response states how many tokens went to input and to output.
3. The gateway multiplies tokens by the model's price and deducts the result from the balance.

So "buy Claude tokens" and "buy Claude credits" mean the same thing here: a balance top-up that can pay for any model in the catalog, not just Claude. One fetch_... key, one balance, models from Claude, GPT, Gemini, DeepSeek, Qwen and GLM. There is no subscription: if you request nothing, nothing is charged.

What a charge consists of

A request has several parts, each with its own price per 1M tokens:

The request cost formula:

cost = fresh input × input price + cache read × read price + cache write × write price + output × output price

The cache is billed separately, and it is seen differently. For Claude the discount comes through the Anthropic format (/v1/messages, Claude Code, the Anthropic SDK): the gateway sees how much input came from cache. In the OpenAI format (/v1/chat/completions) all Claude input is charged at the input price. For GPT models cache reads are visible in both formats, and there is no separate charge for writes. A detailed look: Prompt caching: reads and writes↗.

How many requests $5 covers

Take two request types.

Chat: 2000 input tokens, 500 output tokens. Agent step: 30,000 input tokens (project context), 1000 output tokens, ignoring the cache.

Cost of one request = input × input price + output × output price (price per 1M tokens).

How to read a row, using claude-sonnet-5-5 and chat: 2000 × 0.4 / 1e6 = $0.0008 for input, 500 × 2 / 1e6 = $0.001 for output, total $0.0018. Then $5 / $0.0018 ≈ 2,777 requests. Numbers are rounded down.

The gap between rows is huge: on the same balance you get from about twenty to over a hundred thousand requests. But do not conclude "take the cheapest". Models differ in tasks, and price says nothing about quality on your material. Compare by price, context, intelligence index and on your own examples.

How the cache changes an agent step

The same agent step on claude-sonnet-5-5 in the Anthropic format, if of 30,000 input tokens 25,000 are read from cache, 3000 are fresh and 2000 are written to cache (prices 0.05 for reads and 0.5 for writes):

read: 25,000 × 0.05 / 1e6 = $0.00125 fresh: 3,000 × 0.4 / 1e6 = $0.0012 write: 2,000 × 0.5 / 1e6 = $0.001 output: 1,000 × 2 / 1e6 = $0.002 total: $0.00545 instead of $0.014; $5 covers about 917 steps

This is an illustration, not a promise: the cache share depends on how the tool builds its requests. Check real figures under "Usage" in the dashboard.

"Free AI tokens": what actually exists

There is no free unlimited access: every request costs money. What there is:

• $0.50 for subscribing to the @zukkopro↗ channel. Link Telegram in the dashboard, subscribe and press "Check". The bonus is once per Telegram account.
• The referral program: 10% of invited users' top-ups.

Any "key generators" and "free keys" found online have nothing to do with us. Do not enter your real data on third-party sites for the sake of "free tokens".

How to top up the balance

The minimum top-up is $1. Payment methods:

So for $5 of balance via SBP you pay the equivalent of $5 × 1.085 = $5.425 in rubles at the rate on the payment day. The payment page always shows the final amount to pay. You can top up the balance from Russia by SBP, MasterCard or crypto.

How to avoid unexpected spending

• Limits per key. Each key can have a daily and a monthly dollar limit and a list of allowed models. When the limit is used up, requests get a 429 (key_daily_limit_reached or key_monthly_limit_reached). A separate key with a limit for each project or tool is the simplest protection.
• Watch spend. Balance and charge history are visible in the dashboard ("Usage") and through the API: How to check balance via the API↗.
• When the balance runs out. A request returns 402. Top up and the same key works again.

To start, create a key in the dashboard, top up the minimum amount and make a first request following the quick start↗.

FAQ

Do tokens expire if I do not request anything for a long time?

The balance is charged by requests and there is no subscription fee. We do not guarantee terms for holding a leftover balance here: check the current rules in the dashboard and on the website.

Can I buy Claude tokens specifically?

You top up a shared balance and choose the model in the request: claude-sonnet-5-5, claude-opus-5-5 or any other from the catalog. There is no separate "Claude package".

How many tokens are in one word?

It depends on the language and the model; Russian is usually "more expensive" than English in tokens. The exact number for your text is shown by the token counting endpoint before the request or by the usage field after it.

What if the balance went faster than expected?

Open "Usage" and look at spend by key and model. The most common culprit is a long context: every agent request carries the accumulated history, see Why more tokens were charged↗.