How the maths works
The pay-per-token side is deliberately simple: tokens ÷ 1,000,000 × price per million.
The flat side is just the monthly fee. Where it gets interesting is the break-even volume —
the point where the two lines cross, which is the monthly fee divided by the price per million. Below it,
metered billing wins. Above it, the gap widens forever, because one side keeps climbing and the other doesn't.
We use a single blended price per million rather than separate input and output rates. That keeps the tool honest for any provider: look up your own blended rate, type it in, and the comparison holds. The default of €3.18 is the closed-model reference used on our home page.
Why per-token bills surprise people
Almost everyone underestimates their token usage, and it isn't carelessness — the cost drivers are invisible from where you sit:
- Output costs more than input. Often several times more, so a chatty model is expensive in a way that isn't obvious from the prompt you wrote.
- Every turn resends the whole conversation. A long session pays for the same context again and again. Cost scales with the square of the conversation length, not linearly.
- Failed work still bills. Retries, abandoned tool calls and reasoning tokens are charged whether or not you keep the result.
- Agents multiply everything. A coding agent that reads files, plans, edits and re-checks can burn 10-50x what the same task costs in a chat window.
This is why the number you get from "how much do I type in a day" is usually off by an order of magnitude, and why the first month's invoice tends to be the shocking one.
When a flat rate is the wrong choice
Plenty of the time, honestly. If you're below the break-even volume the calculator shows, pay-per-token is cheaper and you should stay on it — the tool says so outright when that's the case. Metered billing also makes sense for spiky, seasonal workloads, and for anyone who needs a specific closed model that no flat-rate provider can resell.
A flat rate earns its keep when usage is sustained: daily agentic coding, batch pipelines, internal tools that run all week. That's when predictability stops being a nice-to-have and starts being the reason you can plan a budget at all.
What counts as a token
Roughly four characters of English, or about ¾ of a word — so 1M tokens is on the order of 750,000 words. Code tends to tokenize slightly worse than prose because of punctuation and indentation. Every provider counts a little differently, which is another reason to treat any estimate, including this one, as a ballpark rather than a quote.
Common questions
How many tokens does a month of coding use?
It depends far more on how you work than on how much you type. A developer using autocomplete occasionally lands in the tens of millions of tokens a month. Running an agent that reads files, plans and retries pushes you into the hundreds of millions, because every step resends the accumulated context. Agentic tools burn 10-50x what chat-style usage does, which is why bills surprise people who estimate from how much they type.
Why is my LLM API bill higher than I estimated?
Three reasons, usually. Output tokens cost several times more than input tokens. Every turn of a conversation resends the entire history, so a long session pays for the same context repeatedly. And retries, tool calls and reasoning tokens are billed even when the result is discarded. The cost scales with context length and iteration count, not with the size of the answer you keep.
When is a flat monthly rate cheaper than paying per token?
Above the break-even volume, which is simply the monthly fee divided by the blended price per million tokens. Below it, pay-per-token is cheaper and you should stay on it. The calculator shows the break-even point for the numbers you enter, and says so plainly when a flat rate is the worse deal for you.
About llmax.ai
We sell the flat side of this comparison: unlimited tokens on open models at a fixed monthly price, through an OpenAI-compatible API, with zero logs and all processing in the EU. If the numbers above make the case, the plans are here and the FAQ covers the limits, the models and how upgrades work.