You are not paying for your prompt

There is a new page on this site: paste a prompt, pick the model you use, and it tells you what the call costs and what else that money would have bought. Building it turned up three things worth more than the tool itself.

Your prompt is a rounding error

Here is a real prompt — a short code-review instruction, 222 characters, 38 words, about 65 tokens. Ask for a 600-token answer and send it to Claude Opus 5. The bill splits like this:

Ninety-eight percent of that call is the answer. It is not an artefact of this prompt being short: the same split holds for GPT-5.2 at 98.7% and Claude Haiku 4.5 at 97.9%, because output is priced three to five times higher than input almost everywhere, and because answers are longer than questions.

This has a practical edge. People spend real effort trimming system prompts, and on most workloads that effort is aimed at two percent of the invoice. The lever that matters is on the other side: how long an answer you ask for, whether you cap it, and whether the model thinks at length before replying. A reasoning model bills you for tokens you never read, and they land in the expensive column.

That is why the page puts the output box beside the prompt box rather than below it. The number that moves your bill is the one you have to type in yourself.

The spread is three orders of magnitude

The same call across every model that publishes a price, plotted against how each one stands in the comparison table:

That shape is the whole argument. A ranking is a line; this is a plane, and most of the field is not on the frontier of it. Everything above and to the left of a given dot is strictly better value than that dot — cheaper and stronger — and for most models that region is not empty.

Three hundred and forty-six times separates the cheapest call from the dearest. At one call a day nobody cares. At a million calls a month it is the difference between a rounding error and a hiring decision.

Forty-two out of fifty-two are beatable

Fifty-two models in the catalog have both a published price and a standing in the table. For forty-two of them, something cheaper is ranked at least as high.

Claude Haiku 4.5 is the clearest case. It costs $0.00306 for our sample call and sits at 43 overall. DeepSeek V4-Flash costs $0.00018 — seventeen times less — and sits at 60. Sixteen models beat Haiku 4.5 on both axes at once.

That is what the page's first list does: given your model, it shows only what is cheaper and ranked at least as high. Not cheaper. Not better. Both.

Ten models have no better deal available. Claude Opus 5 is one: nothing cheaper matches its standing. That empty list is worth as much as a full one — it means the premium is buying something, and the page says so in words rather than hiding the section.

Below it sits a second list of neighbours in price, a deliberately different question. Those are the models you could switch to without the bill moving, better or worse, and it is where the awkward truths surface: pick Claude Opus 5 and its neighbours are Opus 4.5, 4.6, 4.7 and 4.8 — the same price, lower rankings. Same money, older model. Upgrading costs nothing.

What the counter will not do

It will not tell you the exact token count, and that is a design decision rather than a shortcoming.

A real count needs the tokenizer the model was trained with: a vocabulary of around a hundred thousand learned merges, several megabytes, different for every family. Shipping one to your browser to answer "roughly what will this cost" is the wrong trade. Shipping one and quoting it for all 112 models would be worse — it would look precise while being wrong for most of the table.

So the estimator counts structure instead, using how byte-pair encoders behave rather than what any particular one contains:

On English prose that lands at about four characters per token, which is where the rule of thumb has always been. The page reports a range around it — the sample prompt is "65 tokens, likely 55 to 78" — because the real answer differs by model, and pretending otherwise would be a small lie told very confidently.

A calculator that says "1,247 tokens" for a prompt destined for four different families is wrong three times out of four, and gives the reader no way to know which.

Two things to check before trusting it

The prices are the vendor's standard pay-as-you-go rate, excluding batch discounts, cached-input rates, off-peak windows and priority tiers. If half your traffic is cache hits, your real bill is lower than anything here, and the ranking between models can change with it.

And the overall column is doing a lot of work in the recommendation. It is built from the benchmarks a model's maker chose to publish, discounted for how few of them there are. A model ranked below another can still be the right pick for your job — a coding workload cares about SWE-bench, not a graduate physics quiz. The lists are a starting point for the question, not the answer to it.

The tool is at Token cost. It costs nothing to find out whether you are one of the forty-two.