Free tool · runs in your browser · no signup

One agent is a line item. A swarm is a budget.

Fan out one orchestrator into N sub-agents and the bill multiplies in ways a per-request price never showed you. Pick your models, your fan-out, and your retry rate, and see the cost per run, per day, and per month, priced from Anthropic’s verified list prices. Nothing leaves this page.

These controls start on illustrative values, not a claim about a typical swarm. Measure a few real runs and set your own token averages; your usage dashboard is the ground truth.

Orchestrator workload

The orchestrator reads every sub-agent’s output, so its input side is usually the biggest single context in the run.

5
Sub-agent task size
10%

Share of sub-agent calls that fail validation and rerun once. Swarms retry more than you think; check your logs.

10
Your swarm, priced

Estimated cost per run, your mix

$0.905/run

1 Claude Opus 4.8 orchestrator + 5 Claude Sonnet 5 sub-agents with 10% retries.

A real line item. Worth a caching pass before it grows teeth.

One run, itemized
Orchestrator (Claude Opus 4.8)
75k in / 8k out
$0.575
5 sub-agents (Claude Sonnet 5)
20k in / 2k out each
$0.300
Retries (10% rerun once)
the tax nobody budgets for
$0.030
Per day (10 runs)
$9.05
Per month (30 days)
$272
The surprise factor
10x spread, same run

Your exact run shape, priced end to end on Claude Haiku 4.5: $0.280/run. The same shape on Claude Fable 5.1: $2.80/run. That is a 10x range you control with a config line, before touching a single prompt.

The price table everything derives from
Claude Haiku 4.5$1 in / $5 out per M
Claude Sonnet 5$2 in / $10 out per M
Claude Opus 4.8$5 in / $25 out per M
Claude Opus 5$5 in / $25 out per M
Claude Fable 5.1$10 in / $50 out per M

standard API list prices, verified against Anthropic’s pricing docs on 2026-09-04. models we could not verify are excluded, not guessed.

Before you quote this number
  • Estimates only, at uncached list price. Real runs vary per task.
  • Prompt caching bills cache reads at a steep discount, a tenth of the input rate on most models and as little as a fortieth on Fable 5.1. Swarms that share a big system prompt cache heavily, so real input cost can land far below this.
  • The Batch API is 50% off both directions for work that can wait.
  • Token averages are yours to measure. Log a few runs before trusting any calculator, including this one.
So now what

An estimate is a guess. Your logs are a bill.

This page prices the run shape you told it about. The runs you actually dispatch drift: contexts grow, retries spike, someone quietly switches a role to a pricier model. Want this tracked automatically, per run and per agent, with budget alerts that fire mid-run? That is exactly what we are building: join the Agent Cost Watch waitlist.

And if your spend is per-seat rather than per-run, our Cursor Auto cost calculator does the same honest math for editor traffic.

FAQ

Swarm costs, straight answers.

One run is one orchestrator call plus N sub-agent calls, each role on the model you picked. Cost per call is tokens divided by a million, times the per-million list price for that model. The retry slider reruns that share of sub-agent calls once. Day and month numbers multiply the per-run cost by your runs per day and a 30-day month. Everything runs in your browser, nothing is sent anywhere.
These are the standard per-million list prices verified against Anthropic's pricing docs on 2026-09-04: Claude Haiku 4.5 at $1 input / $5 output, Claude Sonnet 5 at $2 / $10, Claude Opus 4.8 and Opus 5 at $5 / $25, and Claude Fable 5.1 at $10 / $50. Models whose price we could not verify on that page are excluded rather than guessed. Source: https://platform.claude.com/docs/en/about-claude/pricing.
The same run shape, priced end to end on the cheapest table model versus the priciest. With the current table that is Claude Haiku 4.5 versus Claude Fable 5.1, which is exactly 10x on both input and output rates, so the whole-run spread is 10x. The number is computed from the price table for your exact shape, not quoted from folklore. It is the range your model choices move you within before you change a single prompt.
Three big reasons. Prompt caching: cache reads bill at a steep discount to list input price (a tenth on most models, and as little as a fortieth on Fable 5.1), and agent frameworks that reuse a large system prompt cache heavily, which can cut real input cost by a lot. Batch API: non-interactive work queued through the Batch API is 50% off both input and output. And token counts: real agents vary wildly per task, so measure a few runs and put your own averages in. This tool prices standard, uncached, non-batch API calls at list price.