How to use the LLM cost calculator

What the calculator does in one pass

The LLM cost calculator prices a month of tokens on every seller of the models you choose, from published input and output rates. You either pick models yourself or switch to Help me choose, set monthly input tokens, monthly output tokens and the cached share of input, choose a serving tier, and read a table with one row per model and seller, cheapest first. It ranks on price only, because the site holds no quality score.

Every rate in the table carries the page it was read from and the date, and any rate older than 21 days is hidden. A model needs both an input and an output price on record to be costed at all.

Every seller listed sells the model as a reseller or host. Where the maker of a model sells it directly, that will appear as another row; until then the page says so under the table rather than offering a filter whose every answer is the same.

Pick a model, or Help me choose

Pick a model is for when you already know what you want to run. Choose one or more models and the table shows each seller of each. This is where the 93 models with more than one seller earn their place, since the spread between sellers of the same model is the main thing the page shows.

Help me choose is for when you know the workload but not the model. It asks for the longest prompt the model must take and how much you will send, then lists the cheapest models that fit. Models whose context window we do not hold are left out rather than assumed to fit.

Read that list for what it is: the cheapest options that meet a length floor, at list price. It cannot tell you whether any of them answer well for your task. Test your own prompts on the shortlist before choosing.

Setting the sliders

The input and output sliders take monthly totals, not per-request figures. Work them out first: requests per month times average input tokens per request, and the same for output. The monthly estimate method walks through getting those averages from your own logs.

The cached share slider is the fraction of input tokens that repeat a prefix. It applies only to sellers that publish a cached rate; the others are billed at the full input rate, and the table says so under the results. Without a measured share, run the page at zero and again at a high value and treat the two totals as your range.

The settings are kept in the address bar, so a link to the page reproduces the same view for a colleague.

Serving tier and the controls that are greyed out

The tier filter has Any tier, Standard, Batch and Free. Batch listings on record: 121. Free listings: 18. Comparing a batch row with a standard row is comparing a delayed answer with an immediate one, so filter to the tier your workload can actually use.

Data residency and a provisioned tier are drawn but disabled, each with the reason. The sellers priced here do not publish token rates by region, and provisioned capacity is sold by the GPU-hour rather than the token, which the GPU price index covers instead.

Free rows deserve a second look. They may come with tight limits on requests or tokens per day, which the table does not show. A free row that tops the ranking for a production volume is a sign to read the seller's limits before planning around it.

Reading the result without over-reading it

The top row is the cheapest way to run your stated month among the models and tier you selected. Every other row shows how much more it would cost. Change the output slider and watch the order, because sellers with a low input rate and a high output rate move down as output grows.

The page states what the figure leaves out: negotiated discounts are not modelled, so a contracted bill may be lower. Retries, embeddings and other calls you did not count are not in it either. Treat the figure as the list-price floor for the tokens you entered.

When the result surprises you, check the inputs before the rates. A slider set to a weekly figure instead of a monthly one, or input and output totals swapped, produces a confident ranking for a workload you do not have.

The calculator also cannot see the model calls you forgot to count. A month entered without retries, evaluation runs or agent steps gives a total that is low on every row, and the ranking holds only if the missing tokens have the same input and output mix as the ones entered.

Questions the table cannot answer

The table is a ranking of list prices for the tokens you entered. It does not know your rate limits, so the cheapest row may not accept your volume on a new account. It does not know where a seller processes requests, which may rule a row out for a regulated workload.

It also does not know how many tokens a given model will use for your task. Different tokenizers and different answer lengths mean the same job can produce different monthly totals on different models. When comparing across models rather than across sellers of one model, measure the token counts on each first.

For questions about a single model and its sellers, the token price table shows the raw rates and the page each was read from, which is the place to check a figure before quoting it.