.workspace, .details-section, #calculator-startup { display: none; }Enable JavaScript to calculate an estimate. You can still read how the calculator works.

Loading the calculator. If it does not appear, refresh the page or check that JavaScript is enabled.

Estimate model API costs for your AI agent.

One task can trigger many billable model calls. Estimate token costs as context grows, tools return data, and caching changes input prices.

Agent Workload

Choose a usage profile or configure how many sessions and tool calls your agent runs.

Agent runs / day

User messages / run

Tool actions / message

Runtime & caching 22 active days · Cache on

Active days / month Days the agent runs during the month

Pause between messages (min) Longer pauses reduce cache reuse

Parallel tool actions Actions returned in one model request

Tool results kept in context How much tool output remains available

Enable prompt caching Availability and discounts depend on the selected model and provider

Token size assumptions 6 parameters · Profile defaults

Agent instructions (tokens) System prompt and tool definitions

Average user message (tokens) Initial request and follow-ups

Average tool result (tokens) Search, API, browser, or execution output

Tool request output (tokens) Model output used to call a tool

Average final reply (tokens) Response delivered to the user

Reasoning / model request (tokens) Thinking tokens billed as output

Model & Rates

Choose standard model API rates and a market, or enter custom per-million token prices.

Official preset · Tier rules active

USD per 1M tokens

Fresh input Uncached prompt tokens

Cache read Tokens read from prompt cache

Cache write Tokens written to prompt cache

Output Generated and reasoning tokens

Cache & context assumptions Context, TTL, minimum cache size

Input context budget Compaction uses the configured context budget

Cache TTL (min) Provider cache lifetime

Minimum cache size Smallest cacheable prompt

Estimated Monthly Cost

— / month

— per agent run

Model API token costs in USD only. Excludes external tools, hosting, token-hour storage and taxes.

— —

Agent runs —

Model requests —

Peak / run —

With caching —

Without caching —

You save

Token type	Cost	Share
Fresh inputAwaiting calculation	—	—
Cache readAwaiting calculation	—	—
Cache writeAwaiting calculation	—	—
Output + reasoningAwaiting calculation	—	—

—

See how this simulation works

Request-by-request context growth and billed-token audit for one agent run

Context Growth Across One Agent Run

Context tokens Compaction event

Cumulative Billed Tokens Across One Agent Run · Log scale

Fresh input Cache read Cache write Output + reasoning

Cumulative Cost Across One Agent Run

Total Fresh input Cache read Cache write Output + reasoning

Request-Level Execution Trace

Input is the new or rewritten content triggering this request; Context is the full prompt billed by the model. Output includes reasoning.

Turn	#	Request	Input	Response	Output	Context	Fresh input	Cache read	Cache write	Cost

How it works

Why Agent Costs Compound Across a Session

An agent can make several model requests for one user task. Each request can resend the growing conversation, add tool results, and receive a new response. The simulator models that request loop instead of multiplying one prompt by a token price.

The AI Agent Request Loop

1. User Prompt User Goal User asks the agent to research, create, analyze, or complete a task.

2. Model Plan Tool Call Emitted Model reasons and outputs parameters for a search, API, browser, or other tool.

3. Execution Tool Output Collected The environment executes the tool and returns data, an action result, or an error.

4. Resend Context Prompt Rebuilt History and tool results return to the model. Eligible prefixes may be reused from cache.

5. Next Step or Reply Iterate or Resolve Loop repeats until goal is satisfied, then model writes final answer.

Read the full methodology →

Modeling Scope

Key Assumptions

These boundaries matter most when interpreting the estimate.

Generic agent approximation
Models standardized agent tool cycles using representative token averages rather than mirroring any proprietary agent framework verbatim.
Single-model execution
All turns run on the selected primary model rather than routing planning, summarization, or background tasks to cheaper satellite models.
Estimate, not a billing promise
Actual cloud invoices vary with retrieved-content size, cache hit distributions, tool argument verbosity, and vendor rounding policies.

View all assumptions →

Common Questions

AI Agent Cost FAQ

Why does one agent task make multiple model requests?

Planning, parallel tool calls, tool results, evaluation, and the final reply each participate in an iterative request loop.

What is Cache read?

Cache read is a reused prompt prefix billed at the provider's cache-hit rate instead of the ordinary input rate.

Why do some models have no Cache write price?

Some providers populate caches automatically. Their cache misses bill as Fresh input rather than through a separate Cache write meter.

Why does context grow during a run?

The agent resends prior messages and adds tool results or other new content until compaction reduces the retained history.

How accurate is the estimate?

It is a request-level scenario estimate. Actual bills depend on real token sizes, cache behavior, retries, routing, and vendor pricing rules.