Free tool / cost

What will your agent cost?

Type in your traffic and token counts. See the cost per run, per finished task and per month, and which change saves the most.

Open the estimatorHow the maths works

No sign-up / runs in your browser

Plan the bill before it arrives.

Start from a model size

These rates are round example numbers, not any provider’s price list. Replace them with the rates on your own bill.

Traffic

One run = one user task from start to finish.

Planning, tool choice, answer, checks. Count them all.

%

Failed runs still cost money.

%

Timeouts, bad tool output, user retries.

Tokens per call

System prompt, history, retrieved text and tool results.

What the model writes back, including tool call arguments.

%

The stable start of your prompt, if your provider caches it.

Prices (USD per million tokens)

Copy it from your provider’s price page.

Output is usually several times the input price.

%

How much cheaper cached input is. Zero if unsure.

Search APIs, vector queries, code sandboxes.

Per month

$1,788

$59.62 a day, 30 day month

Per finished task

$0.0324

Spend divided by runs that actually succeed

Per run

$0.0298

$0.0276 before retries

Spent on failures

$143

Plus $132 a month on retries

Where you could save

Each idea is tested on its own against your numbers above. They do not add up exactly.

  • Send half the calls to a small model

    −$760 /mo

    Only safe if quality holds on your own evals.

  • Cache 70% of the input

    −$735 /mo

    Keep the stable prefix of the prompt identical between calls.

  • Trim 25% of the input tokens

    −$292 /mo

    Shorter prompts, fewer retrieved chunks, summarised history.

  • Halve the retry rate

    −$66.24 /mo

    Bounded retries and better tool error messages.

This is a planning estimate from the numbers you typed. To see what your agent really spends, run by run, send a trace to Tracify and compare.

Read next

Questions people ask.

How is the monthly cost worked out?

Each model call costs its input tokens times the input price plus its output tokens times the output price. That is multiplied by the calls per run, repeated for the share of runs that retry, then multiplied by runs per day and 30 days.

Where do the example prices come from?

They are round numbers for a small, a mid and a frontier model, so the page works out of the box. They are not a provider price list. Type the rates from your own provider bill.

Why is cost per finished task higher than cost per run?

Runs that fail still spend tokens. Dividing the total by only the runs that succeed shows what a delivered result really costs.

Does this page store my numbers?

No. The maths runs in your browser. The copy link button puts your numbers in the web address so you can share them, and nothing is saved on our side.

How accurate is it?

It is as accurate as your inputs. Real agents vary run to run, so treat the result as a planning range and check it against measured spend once you have traces.

Now measure the real thing.

Send a first trace and see the true cost of every run.

Start freeRead the quickstart