CheapAI
12 questions

Questions worth answering properly

Model authenticity, where the savings come from, limits, billing and the errors people actually hit.

01Is this the same model the vendor serves?

Yes — same model ids, same API shapes, same capabilities. response.model echoes exactly what you requested, so you can verify it on any call. Nothing is silently swapped for a cheaper model.

02How can it be cheaper than going direct?

Volume purchasing upstream plus a thin margin on top. There is no subscription and no monthly minimum — a prepaid balance is drawn down per token.

03Do I need a VPN?

No. The gateway handles the upstream side, so it works from wherever you are.

04Which models are available?

Claude (Opus, Sonnet, Haiku, Fable), the GPT-5.x line including Codex variants, Gemini 3.x Pro and Flash, Grok 4.x, plus image models billed per request. The live catalog is always current.

05Does prompt caching work?

Yes, on Claude models, billed at the reduced cache-read rate. There is a runnable example.

06Is web search billed separately?

No. Search and grounding results bill as ordinary input tokens. Anthropic charges $10 per 1,000 web_search invocations on its own API and Google bills grounded requests separately — here it is just tokens.

07How do I get citations out of Grok?

Use /v1/responses with tools: [{"type": "web_search"}]. Structured url_citation annotations only come back on that endpoint — /v1/chat/completions will search but strips the citation array.

08Why is my output capped at 1024 tokens?

That is an SDK default, not a gateway limit. The Anthropic SDK sets max_tokens: 1024 when you omit it. Pass your own value.

09My stream cuts off partway through.

Raise the client read timeout to 120–180 seconds — reasoning models can stay quiet for a long time before the first token. If the same request succeeds through curl but fails in your tool, the problem is on the client machine: a proxy, a TLS-inspecting antivirus, or an idle timeout.

10What are the rate limits?

Up to 500 requests per minute per key. There is no fixed concurrency cap — throughput is bounded by upstream capacity.

11How do I pay?

Rubles via SBP, card or SberPay; USDT on TRC20, BEP20, Polygon or TON; USDC, BTC, ETH, LTC, TRX and SOL; card and PayPal internationally. From $5 — see payment options.

12Is there a free tier?

New accounts get a small starting balance to try things. After that it is prepaid, from $5. Create an account.

Still stuck? The documentation goes deeper, and TROUBLESHOOTING.md maps symptoms to causes.

One key. Every frontier model. A fraction of the price.

Pay in rubles, USDT or by card. No VPN, no subscription, from $5.

Get an API key