Why is one agent task not one API call?

An agent works in a loop: it reads the task, decides to use a tool, calls it, reads the result, and thinks again. Each of those turns is a separate request to the model. A support agent that looks up an order, checks a refund policy and drafts a reply has already made three or four calls — and each one is billed in full.

Why does a longer agent run cost so much more than a longer chat?

Because the conversation is resent every step. Step eight carries the system prompt, the tool definitions, the original task, and everything the first seven steps said and found. So input tokens grow with roughly the square of the number of steps, not in proportion to it. Doubling an agent from five steps to ten more than triples its input, and that is the single thing most cost estimates miss.

How many model calls should I put per task?

Count the tools your agent can reach for and assume it uses most of them once, plus one call to start and one to write the final answer. Then check a real run: almost every framework logs the number of steps. The calculator starts from typical figures per kind of agent, and they are meant to be replaced by your own.

Does prompt caching really help that much?

More than anywhere else. Every step of an agent resends the same system prompt and tool definitions, and a cache read costs about a tenth of the input price. On a long run with a big toolset, caching the repeated prefix routinely takes half the bill off. See [prompt caching](https://platform.claude.com/docs/en/build-with-claude/prompt-caching) for how the discount is applied.

Do the tool definitions themselves cost anything?

Yes, on every single call. The tools you declare are sent as part of the request, and the provider also adds a system prompt to enable tool use — Anthropic publishes the exact token count per model, in the high hundreds. Ten tools with detailed schemas can be a couple of thousand tokens resent every step, which is why the calculator asks for the system prompt and tool definitions as one number. See [tool use](https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview).

What about steps that fail and get retried?

They are billed like any other call, and they cost more than average — a retry happens partway through a run, so the context it resends is already long. The calculator takes a retry rate and adds whole calls at the end of the loop, where they are most expensive, rather than averaging them away.

Is the cost per task or per call the number I should watch?

Per task. It is the unit your business actually produces — a resolved ticket, an extracted document, a booked appointment — and it is what you compare against what the task is worth or what a person doing it would cost. Cost per call is only useful for spotting that a loop is running longer than you thought.

How can I make an agent cheaper without making it worse?

In order of effect: cache the repeated prefix; return less from your tools, since a search result page inflates every later step; cap the number of steps; ask for shorter reasoning, because output costs several times what input does; and use a cheaper model for the routing steps and an expensive one only for the final answer.

Are these prices current?

Every rate comes from the provider's own pricing page and is kept up to date — [OpenAI](https://developers.openai.com/api/docs/pricing) and [Anthropic](https://platform.claude.com/docs/en/about-claude/pricing) among them. If a price has changed, or you have negotiated a discount, enter your own under Advanced settings.

AI Agent Cost Calculator

What One Agent Task Really Costs

Inputs

Which model does it run on?

Model

What does your agent do?

Closest to your agent

Model calls per task

This is the number people get most wrong. An agent that looks something up, checks a rule and then writes a reply has already made three or four calls, not one.

How much work does it do?

Tasks

Per

Not sure about the numbers?

Two ways to measure one step of your agent

Cache the system prompt and toolsGrows each stepStandard0% retries

Results

Cost per task

$0.00

One task · 5 calls to GPT-5.6 Terra

Cost per model call

$0.008640

averaged across 5 calls

Context growth

3.3×

1,800 tokens on the first call, 6,000 on the last

What drives one task

$0.0432

Standard input$0.0270

Cached input$0.001200

Model output$0.0150

At this volume

100 tasks · Day

Per day

$4.32

Per month

$131.40

Per year

$1,576.80

If you ran this many tasks

100

$4.32

1,000

$43.20

10,000

$432.00

100,000

$4,320.00

What this figure assumes

5 calls3.3× context20% saved by cache

Model calls per task: 5 calls · 0% retries
Context: Grows each step · 3.3× growth
tokens: 19,500 input · 1,250 output
Prompt caching: Cache the system prompt and tools · 6,000 tokens
Model: GPT-5.6 Terra · $2 in / $12 out

Prompt caching saves 20% · $0.0108 per task

Is $131.40 a month worth it? Work out the ROI

Beyond this calculator

Upload instead of typing
Drop in a paystub, bank statement or screenshot — the AI reads it and fills these fields for you.
Keep the result
Income & Expense Tracker is where this stops being a one-off number: keep your real balances there and the projection updates itself.
Documents in one sentence
Describe an invoice and the AI writes it. PDF without watermarks, free.

AI Spend Desk

An agent is priced per task, and the task is not one request

The number that decides an agent's economics is not the price per token — it is how many times the loop turns, and how much conversation each turn has to carry. A five-step agent resends its system prompt five times and its growing history four times on top; a fifteen-step agent does something far worse than three times that. Rates here come from the providers' own pricing pages, OpenAI and Anthropic among them.

Count the steps from a real run

Almost every agent framework logs the number of model calls per run. One real trace beats any estimate, and it is usually higher than the number people guess.

Cache the repeated prefix

The system prompt and tool definitions are resent on every step. At about a tenth of the input rate, caching them is the single biggest saving available to an agent.

Watch what your tools return

A tool that hands back a page of JSON does not cost you once — it costs you on that step and on every step afterwards, because it stays in the conversation.

Compare cost per task with the task’s worth

A resolved ticket, an extracted invoice, a booked slot. If the agent costs less than the outcome is worth, the token price stops being the interesting question.

Price the loop, not the request. Then cache the part that repeats, trim what the tools return, and cap the steps — in that order.

Frequently Asked Questions

An agent works in a loop: it reads the task, decides to use a tool, calls it, reads the result, and thinks again. Each of those turns is a separate request to the model. A support agent that looks up an order, checks a refund policy and drafts a reply has already made three or four calls — and each one is billed in full.

Related calculators

AI Agent Cost Calculator

What One Agent Task Really Costs

Inputs

Which model does it run on?

Model

What does your agent do?

Closest to your agent

Model calls per task

This is the number people get most wrong. An agent that looks something up, checks a rule and then writes a reply has already made three or four calls, not one.

How much work does it do?

Tasks

Per

Not sure about the numbers?

Two ways to measure one step of your agent

Cache the system prompt and toolsGrows each stepStandard0% retries

Results

Cost per task

$0.00

One task · 5 calls to GPT-5.6 Terra

Cost per model call

$0.008640

averaged across 5 calls

Context growth

3.3×

1,800 tokens on the first call, 6,000 on the last

What drives one task

$0.0432

Standard input$0.0270

Cached input$0.001200

Model output$0.0150

At this volume

100 tasks · Day

Per day

$4.32

Per month

$131.40

Per year

$1,576.80

If you ran this many tasks

100

$4.32

1,000

$43.20

10,000

$432.00

100,000

$4,320.00

What this figure assumes

5 calls3.3× context20% saved by cache

Model calls per task: 5 calls · 0% retries
Context: Grows each step · 3.3× growth
tokens: 19,500 input · 1,250 output
Prompt caching: Cache the system prompt and tools · 6,000 tokens
Model: GPT-5.6 Terra · $2 in / $12 out

Prompt caching saves 20% · $0.0108 per task

Is $131.40 a month worth it? Work out the ROI

Beyond this calculator

Upload instead of typing
Drop in a paystub, bank statement or screenshot — the AI reads it and fills these fields for you.
Keep the result
Income & Expense Tracker is where this stops being a one-off number: keep your real balances there and the projection updates itself.
Documents in one sentence
Describe an invoice and the AI writes it. PDF without watermarks, free.

AI Spend Desk

An agent is priced per task, and the task is not one request

Count the steps from a real run

Almost every agent framework logs the number of model calls per run. One real trace beats any estimate, and it is usually higher than the number people guess.

Cache the repeated prefix

The system prompt and tool definitions are resent on every step. At about a tenth of the input rate, caching them is the single biggest saving available to an agent.

Watch what your tools return

A tool that hands back a page of JSON does not cost you once — it costs you on that step and on every step afterwards, because it stays in the conversation.

Compare cost per task with the task’s worth

A resolved ticket, an extracted invoice, a booked slot. If the agent costs less than the outcome is worth, the token price stops being the interesting question.

Price the loop, not the request. Then cache the part that repeats, trim what the tools return, and cap the steps — in that order.

Frequently Asked Questions

An agent works in a loop: it reads the task, decides to use a tool, calls it, reads the result, and thinks again. Each of those turns is a separate request to the model. A support agent that looks up an order, checks a refund policy and drafts a reply has already made three or four calls — and each one is billed in full.

AI Agent Cost Calculator

Results

$0.00

Beyond this calculator

Upload instead of typing

Keep the result

Documents in one sentence

An agent is priced per task, and the task is not one request

Count the steps from a real run

Cache the repeated prefix

Watch what your tools return

Compare cost per task with the task’s worth

Frequently Asked Questions

Agent cost by use case

Guides

Related calculators

AI Subscription vs API Calculator

AI Token Counter & Cost Calculator

AI Image Generation Cost Calculator

AI ROI Calculator

AI Cost Calculator

Trading Journal

AI Agent Cost Calculator

Results

$0.00

Beyond this calculator

Upload instead of typing

Keep the result

Documents in one sentence

An agent is priced per task, and the task is not one request

Count the steps from a real run

Cache the repeated prefix

Watch what your tools return

Compare cost per task with the task’s worth

Frequently Asked Questions

Agent cost by use case

Guides

Related calculators

AI Subscription vs API Calculator

AI Token Counter & Cost Calculator

AI Image Generation Cost Calculator

AI ROI Calculator

AI Cost Calculator

Trading Journal