Why the price through a gateway is lower than official — how it works

It's a fair question: if the same models sit under the hood — claude-opus-5-5, gpt-5.6-sol, and so on — where does the difference in your final bill come from compared to an official subscription or the official API directly? The answer isn't a secret discount negotiated with the model provider — it's a handful of concrete mechanisms visible right in the API's own documentation. The per-token rates themselves are also below official pricing (for example, claude-opus-5-5 is $0.30/$0.30 per 1M tokens versus the vendor's $4/$20), with one caveat: vendors charge about 10% of the input price for prompt-cache reads, while here cached tokens are billed at the normal input rate, so for agents with a large repeated context the real gap is smaller than a head-to-head comparison suggests.

Paying per token, not for a subscription with a ceiling

Official subscription tiers for model usage are often sold as a bundle: a fixed monthly amount for access, regardless of how many tokens actually get used. If usage is uneven — a busy week followed by a quiet stretch — overpaying for idle time is baked into the subscription model itself.

Here, billing runs against a prepaid balance for tokens actually consumed, not against a calendar period. The balance is charged exactly the amount of the request — visible right in the X-Zukko-Cost-Micros header (cost in microdollars, 1,000,000 = $1) of every regular (non-streaming) response, with the remaining balance after the charge in X-Fetch-Balance-Cents (despite the name, also in microdollars). Streaming responses don't carry these headers — their spend shows up in the dashboard and /v1/usage. Idle time between requests costs nothing, because there's simply nothing to charge.

Balance checked before the request, not after

There's a less obvious source of savings: the balance is checked before the request ever reaches the model. If the balance is insufficient, you get a 402 instantly:

{"error": {"message": "Недостаточно средств на балансе", "code": "insufficient_balance"}}

A request that fails the balance check never reaches the model provider and costs nothing. In schemes where charging happens after the fact or the check runs asynchronously, that guarantee doesn't always hold — here it's built into the architecture.

Response caching removes exact-repeat charges entirely

A separate, real source of savings isn't a pricing policy but an engineering mechanism: with the X-Zukko-Cache: true header (temperature: 0 also enables it, but only on models that accept it — Claude models may reject a temperature other than 1), identical requests from the same user get cached (up to 7 days for chat). A repeated identical request comes back instantly and free, with an X-Zukko-Cache: hit header in the response. For scenarios like repeated system-level queries, automated tests with fixed prompts, or frequently repeated user questions, this noticeably cuts the total bill with no effort beyond turning caching on.

Idempotency-Key removes accidental double charges

Idempotency-Key is a header that guarantees a repeat request with the same key within 24 hours returns the same response without charging again. At first glance this looks like protection against client bugs rather than a savings mechanism — but in practice, repeated calls from network drops and retries add up to a real amount at meaningful traffic volume, and here that amount simply never appears.

Local payment rails instead of an international card

A separate cost that usually doesn't show up in a "price per token vs. price per token" comparison is the cost of the payment itself. Official payment methods often require an international card, carrying an exchange-rate markup and a currency-conversion fee from the card issuer — a cost that isn't part of the model's price, but is very much part of what you actually pay.

Here, available methods include SBP, a bank card via PayPear, crypto, and — starting 2026-09-16 — a foreign Mastercard EU card via Fride (that one carries a fee of about 13%, amounts $2–$594). None of this changes the model's own price, but it removes the payment-path overhead that, over a month of typical usage, sometimes turns out comparable to the token spend itself at lower volumes.

What isn't part of this savings story

It's worth not confusing these mechanisms with myths. There is no starting balance on regular signup — that practice was discontinued on 2026-09-18 because of abuse from disposable-email farming, so don't count on a free dollar at signup. The one guaranteed bonus is a one-time $0.50 for subscribing to the Telegram channel @zukkopro↗, which requires linking Telegram in the dashboard, subscribing, and pressing "Check"; it's given once per Telegram account.

There's also a referral program — 10% of top-ups made by users you refer — but that's a mechanism for people who bring in others, not a blanket discount on every request.

The bottom line, in mechanisms rather than slogans

Put together: pay-as-you-go billing with no minimum monthly commitment, free caching on repeated requests, protection against duplicate charges via Idempotency-Key, and local payment methods (SBP and Russian cards, charged in rubles at the market rate) — these are the specific places where the final bill ends up lower than a direct official subscription with fixed pricing paid by international card. None of this requires taking anything on faith — the cost of each request and the remaining balance are right there in the API's response headers, and the full history sits behind GET /v1/usage.