$5 Today, Zero Monthly Fee: The Pre-Top-Up Checklist

01 / What a $5 top-up has to cover before the first request

Start with the smallest amount the site will accept as a top-up. If that is 5 dollars, treat it as a hard ceiling for the first month rather than a starting point you will grow. A tiny budget only works if the first experiment cannot run away from you, so fix the number before you look at any model list or any rate table. Write it down next to the date, because a balance with no date attached quietly becomes a balance you stopped watching.

Then separate the two pricing shapes you will meet. A fixed monthly fee is owed whether you call the API or not: a 20 dollar monthly charge is 240 dollars a year of guaranteed spend, including the weeks where you sent zero requests. Pay-as-you-go billing with no monthly fee and no subscription inverts that risk, because the spend stops the moment you stop sending. CheapRelay is built on that second shape, and one key there reaches Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok and more, so you do not pay twice just to try two ecosystems.

The honest question is not which relay has the lowest headline rate. It is how much you can lose if you misunderstand the pricing. With a subscription, the answer is the full monthly fee plus whatever overage is billed on top. With a pay-as-you-go relay, the answer is the top-up you loaded and nothing else, which is exactly the property a beginner should buy first.

02 / Tokens you pay for include the ones you never see

Per-token billing looks simple until caching enters the picture. Some providers quote a low cached-input rate and then add a cache write charge on top of it. CheapRelay bills per real token, and that includes cache reads and writes, so the lines on your receipt add up to the tokens actually processed instead of to a surprise multiplier at the end of the month. That single rule removes most of the accounting anxiety a beginner has.

Do the arithmetic on one real session before you top up. Suppose your system prompt is 12,000 tokens and your question is 3,000 tokens. The first call bills 15,000 input tokens plus the answer. Run that call ten times in one evening and you have billed 150,000 input tokens before any output is counted. At a rate of r dollars per million tokens, the input side alone costs 0.15 times r. Whether that is pocket change or your entire budget depends on r, which is why the rate is checked before the top-up and not after.

Three questions settle most surprises: are cache reads billed and at what rate, are cache writes billed, and is the quote per million input tokens and per million output tokens or blended into one number. Get those three numbers in writing on the pricing page. If a relay hides them, assume the expensive interpretation until the first receipt proves otherwise.

03 / The base URL decides whether your tools work tonight

CheapRelay exposes an API base at api.cheapapirelay.com, with OpenAI-compatible routes under /v1 and Gemini native routes under /v1beta. That split matters more than it looks. Tools that speak the OpenAI dialect, including anything built on an OpenAI SDK, talk to the /v1 routes, while Gemini-native clients expect the /v1beta shape. Guessing wrong costs you a debugging evening and a few thousand billed tokens of failed retries before you learn anything.

Before you fund anything, confirm the tool you actually use appears in the compatibility list. Claude Code, Codex and Gemini CLI are supported, and so is any OpenAI SDK client. Check the same three items for your own setup: the base URL, the path prefix, and the model string your client sends on each call. A relay can be cheap per token and still burn your budget if every request fails authentication and the client retries it three times.

04 / Price two model names before you commit to a long list

A long model catalogue is a feature for someone with a budget and a trap for someone without one. CheapRelay reaches a wide set of families, but you only need two numbers to begin. Choose one cheap general model for the bulk of your work and one stronger model for the occasional hard task, then look up both rates side by side.

On sale at the moment are DeepSeek V4 Flash and Claude Haiku 4.5, and those two make a sensible beginner pair: a fast low-cost model for everyday calls and a higher-quality model for the calls where quality actually changes the outcome. Note the word pair. A beginner who prices eight models ends up choosing none of them and sending mixed traffic that is impossible to read on a receipt.

Write the split down as a rule before you start. For example, run everything on the cheap model by default and switch to the stronger model only when the cheap model fails the same question twice. Then count calls rather than impressions. Ten cheap calls and one strong call is a bill you can predict. Fifty mixed calls is a bill you can only read afterwards.

05 / Fine print that decides whether the $5 lasts a week or a month

Before the payment screen, read the boring pages. They are where a small budget is either protected or quietly drained, and none of the items below need a benchmark or a sales claim to check.

• ■ Minimum top-up and whether smaller test amounts are allowed at all.
• ■ Balance expiry, so a 5 dollar credit does not vanish while you are learning.
• ■ Refund policy on unused credit if the relay turns out not to fit your tooling.
• ■ Rate limits and what a burst of retries does to your balance.
• ■ Logging and data retention, so your prompts are not stored somewhere you did not agree to.

CheapRelay keeps the billing model deliberately thin here: per real token, including cache reads and writes, no monthly fee and no subscription. That means the only variables left are how many tokens you send and which model answers them, both of which are under your control.

06 / A five-dollar smoke test with a hard stop

Treat the first day as an experiment with a receipt, not as onboarding. The goal is to convert an unknown balance into four known numbers: cost per small call, cost per cached call, cost per long call, and cost per strong-model call. Everything after that is arithmetic.

• ■ Send one short call, roughly 2,000 input tokens and 200 output tokens, and note the exact token counts returned.
• ■ Send the identical call again and compare the balance delta to see whether caching changed the price.
• ■ Send one long call with a 12,000 token context and record the difference.
• ■ Send one call to the stronger model and record its rate separately.
• ■ Multiply each delta by a realistic weekly volume before you decide whether the budget is enough.

If a session runs long, the hard stop is the balance itself: when it is gone, you stop and think instead of refilling by reflex. Keep the four numbers in a note file, and add a line for every week you spend. After two weeks you will know whether you can afford the stronger model on a deadline and whether a bigger top-up is worth it.

One last lever is worth knowing about before you spend anything: the referral reward pays 20 percent of what invited users pay, which is a small but real offset for a beginner who learns alongside friends. If the checks above pass, sign up at cheapapirelay.com, load the 5 dollars you decided on, and let the first receipt teach you the rest.