$3.80 a Month After Dropping the Free Tiers

01 / Free tiers ran out on request 61 of the day

My hobby project is a small reading assistant that summarizes articles I save. Nothing fancy, no customers, just me and a laptop. Last spring it made about 1,200 requests a month, roughly 40 a day, and each request ran between 2,000 and 3,000 tokens once you count the prompt, the saved article, and the reply. That is about 3 million tokens a month. Free tiers stopped absorbing that quietly. I hit hourly rate limits mid-afternoon, got hard monthly caps with no warning, and ended up juggling two accounts and three environment variables just to keep a hobby tool alive. The real cost was not money, it was the guessing: I never knew whether the next call would work.

So I moved to pay-as-you-go on purpose. The goal was not performance or prestige, it was a predictable number at the end of the month. I set a ceiling first and only then looked at providers, which is the opposite of how most people do this.

02 / A five-dollar ceiling before a single line of code

Before touching any API, I wrote down four rules on paper. They cost nothing and they saved me from the usual overspend spiral.

• ■ Hard monthly cap of 5 US dollars. If the account reaches it, the tool stops. No exceptions for a weekend project.
• ■ One test load of 1 dollar to prove the integration works before I trust it with real traffic.
• ■ A daily self-imposed limit of about 20 cents, tracked in a tiny text file the script appends to.
• ■ One alert at 50 percent of the cap, so I get a warning at 2.50 dollars rather than a surprise at the end.

Those numbers are arbitrary but they are mine, and that is the point. With a free tier you have no budget to enforce because the price is zero until it suddenly is not. With metered billing you can decide in advance exactly how much a hobby is worth to you. Five dollars a month is less than a coffee, and it is enough to run a small tool thousands of times if I am not careless.

03 / One key and one base URL replaced four signup forms

The setup part took about fifteen minutes. I signed up at CheapRelay, loaded 1 dollar, and got one API key. CheapRelay is a low-priced pay-as-you-go relay where a single key reaches Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok and more, and it works with Claude Code, Codex, Gemini CLI, and any OpenAI SDK.

The configuration is deliberately boring:

• ■ API base URL: https://api.cheapapirelay.com
• ■ OpenAI-compatible routes live under /v1, so my existing SDK code needed only the base URL swapped.
• ■ Gemini native routes live under /v1beta if you prefer that wire format.
• ■ One environment variable for the key, and nothing else in my config.

My old setup had four provider accounts, four keys, and a hand-rolled fallback function that guessed which service would fail next. That whole layer disappeared. I now call one base URL and pick the model name per request, which means I can change brains without changing plumbing. For a beginner, that is the underrated win: fewer moving parts means fewer 2 AM debugging sessions.

04 / Token math that includes cache reads and cache writes

Billing is per real token, including cache reads and writes, with no monthly fee and no subscription. Read that twice, because it changes how you build. Prompt caching is a genuine saving when you resend a large stable block, such as my 8,000 character system prompt, on every call. But the cache write is billed too, and cheap models are sometimes not worth caching at all if the write costs more than the read saves.

So I count three numbers per request instead of one: input tokens, output tokens, and cache tokens. My rough monthly shape at 1,200 requests looks like this: about 2.1 million input tokens, about 600,000 output tokens, and a few hundred thousand cache tokens of both kinds. Output is the expensive direction on most models, so I cap answers at 400 tokens for routine summaries. That single line cut my output volume by more than half compared with my first messy draft. I will not quote you a per-token price here, because rate cards change; read the current one before you commit, and multiply your own usage against it.

05 / Cheap models carry the ninety percent of boring calls

Routing is where a tiny budget is won or lost. I keep one strong model for the hard 10 percent, the long documents and the ambiguous queries, and a fast cheap model for everything else. Two models currently on sale fit my split nicely: Claude Haiku 4.5 for the high-volume routine summaries, and a stronger model when I genuinely need deeper reasoning.

• ■ Roughly 1,080 requests a month go to the cheap model.
• ■ Roughly 120 requests a month go to the stronger model, only when I ask for it explicitly.
• ■ Failures are retried on the cheap model first, so a retry does not double my spend.
• ■ I log the model name next to each entry, so I can see which calls earned their cost.

That discipline is what turns a 5 dollar ceiling into a comfortable margin instead of a tightrope. Notice I never run a stronger model by default just because it exists. A hobby project does not need the best answer every time, it needs a good enough answer every time for pennies.

06 / The month-one receipt line and what I changed

At the end of my first metered month the log summed to 3.80 dollars: about 2.20 dollars of input, about 1.30 dollars of output, and about 0.30 dollars of cache traffic. That is the receipt line, and it is the number I now plan around. Under my 5 dollar ceiling with room to spare, no rate limits at 2 PM, no second account, no guessing. The 1 dollar test load was still sitting there unused as margin.

Two changes came out of it. First, I trimmed my system prompt from 8,000 characters to 3,000, because a shorter prompt is cheaper than a cached long one for my volume. Second, I added a weekly check of the token log so drift shows up early instead of at month end. Total time spent on cost management: about ten minutes a month.

If you are a beginner moving off free tiers, the sequence matters more than the provider: set a hard cap, load a small amount, wire one base URL, count cache tokens honestly, and route the boring 90 percent to a cheap model. CheapRelay is the relay I use because the billing is metered per real token with no monthly fee, and it takes the multi-provider juggling out of the picture. Sign up at CheapRelay, load a dollar, and run your own month-one receipt line before you commit to anything bigger.

One last practical note: there is a referral reward of 20 percent of what invited users pay, so if you do bring a friend along, the credit quietly softens your own bill. I would rather you spend 3.80 dollars of real usage and know exactly where it went than chase free tiers that vanish mid-project.