AI Agent Cost Calculator

Agents loop: think, call a tool, read the result, think again. Each step is a billed model call, so cost scales with steps rather than tasks. This tool prices a full agent run.

Inputs

Result

$594

per month

$0.0990 per task · $19.80 per day

Model calls per task
8
Input tokens per task
24,800
Output tokens per task
2,800
Cost per step
$0.0124
Cost per task (incl. retries)
$0.0990
Cost per 1,000 tasks
$99.00
Cost per day
$19.80
Cost per month (30 days)
$594

Disclaimer: Model availability, pricing and specifications may change. Verify pricing directly with the provider before making business decisions. Preset rates reflect published list prices last reviewed August 2026; choose “Custom pricing” to enter your own rates.

Formula

  • Step input = prompt tokens + tool result tokens
  • Task cost = (step input × steps × input rate + step output × steps × output rate) ÷ 1,000,000
  • Task cost is then multiplied by (1 + retry rate) and by tasks per day

Methodology

An agent is a loop, not a single call. Each iteration sends the running scratchpad plus the latest tool output and receives a short decision, so the input side dominates and total cost is roughly proportional to the number of steps the agent takes.

Tool results are billed as input the moment they enter the context, which is why a verbose API response or a full web page can cost more than the model's own reasoning.

Real agents fail and retry. The retry percentage adds that overhead so the estimate reflects production behaviour rather than the happy path.

Example: an eight-step research agent

  1. Each step sends 2,500 prompt tokens plus a 600-token tool result and returns 350 tokens, across 8 steps, with a 10% retry rate.
  2. Input: 3,100 × 8 = 24,800 tokens → $0.062. Output: 350 × 8 = 2,800 tokens → $0.028.
  3. With retries: ($0.062 + $0.028) × 1.1 = about $0.099 per task, roughly $19.80 a day at 200 tasks and about $594 a month.

Frequently asked questions

Why are agents so much more expensive than a single prompt?

One user request becomes many model calls, and each call resends the accumulated scratchpad. Ten steps can easily cost twenty times a single completion.

What is the biggest lever on agent cost?

The step count. Constraining the loop with a maximum-step limit, better tool descriptions and early stopping cuts cost faster than any pricing change.

Should tool outputs be truncated?

Almost always. Summarise or filter API responses before they enter the context; raw payloads are frequently the largest input component in an agent run.

Are these prices always up to date?

The presets reflect widely published list prices and are reviewed periodically, but providers adjust them and offer discounts for batch processing, cached input and committed spend. Model availability, pricing and specifications may change. Verify pricing directly with the provider before making business decisions.

What happens when a provider launches a new model?

Model names and rates live in one central configuration, so a new model appears in every relevant dropdown as soon as it is added. Page URLs stay provider-based rather than version-based, so no link ever goes stale when a model generation changes.

Does the calculator send my data anywhere?

No. Every calculation runs in your browser, so nothing you type is uploaded, stored or logged.

Related AI calculators