1. Price exclusive token buckets
Input = initial input + growth × loop index. Split it into ordinary, cache-read and cache-write buckets, then apply each rate. Add output tokens and per-attempt fees.
PLAN THE WORK. PRICE THE LOOPS.
Agent Budget Planner.
Know what one workflow costs before you run it. Explore cache misses, longer context and retries without spending API credits.
Editable JSON only · maximum 256 KiB · version 1.0.1. A failed import leaves your current plan intact.
Ready. Enter your own rates or load the fictional sample.
01
Set the planning frame
Use one consistent currency. Changing its label does not convert any amounts.
02
Model the calls in each loop
Count all fan-out and subagent calls here. Different models or rates belong in separate stages. Up to 12 stages.
03
Compare two scenarios
Normal
Stress
0.1 means 1.1 billed attempts per logical call on average. Retries use the same entered tokens and fees. Stress uses the entered retry cap for every call; these are planning assumptions.
An entered cap must cover both scenarios. It is a reviewed assumption, never an installed limit or guaranteed cost ceiling.
04
Optional: budget burn at full occupancy
Slots change burn rate, not per-workflow cost. This assumes continuous occupancy with no queueing or rate-limit simulation. Fan-out is already counted in stage calls.
See the cost of your assumptions.
Stress is a conditional estimate, not a worst-case bound. Actual usage and external or in-flight charges can exceed it.
Fix the highlighted input to continue.
Results and exports are hidden until all inputs are valid. Select the error to focus the field.
Zero rates or zero activity can produce zero estimates. This does not imply your provider is free.
Per-workflow breakdown
| Per-workflow breakdown | Normal | Stress |
|---|
Per-stage detail
Take the whole plan with you.
JSON includes inputs, assumptions, source links and exact rational results. CSV prefixes formula-like labels; spreadsheet import behavior varies, so review untrusted text.
Assumptions worth checking
Before running your agent
Review these in your real runtime. Checking a box here installs or enforces nothing.
A transparent calculation, not a black box.
2. Count all billed attempts
Multiply by calls per loop and 1 + extra attempts. Sum every loop and stage, then add the fixed workflow fee once. Input growth is counted on every successive loop.
3. Scale activity separately
Period cost = workflow cost × starts/day × days. Burn/hour = workflow cost × occupied slots × 3,600 ÷ seconds/run. Budget-fit counts use exact arithmetic.
What should I include in token counts?
Include billed system prompts, tool definitions, conversation history and retrieved content. Include billed reasoning in output when your provider does, without double counting. Use a separate stage or plan for different context price bands.
Are these current model prices?
No price catalog is bundled. Enter applicable API prices from your provider or contract. The sample is clearly fictional. API prices are not subscription-plan allowances, and currency labels never convert amounts.
What are the limits and precision?
Up to 12 stages and a 256 KiB JSON import. Integer loops, days and calls up to 1 million; token counts up to 1 billion; extra retries up to 10,000. Prices up to 1 million per million tokens, fees up to 1 billion, budgets up to 1 quadrillion. Decimal fields accept up to 9 places. Exact fractions drive calculations and budget floors; display values are rounded only for readability.
Use your applicable provider prices
Check contract terms, service tier, region, context thresholds, cache TTL, taxes and tool charges. This tool applies one entered rate set per stage.