Should I fan out?

Splitting a task across subagents feels like it should make it proportionally faster at about the same cost. It doesn’t. Serial work caps the speedup, integration is serial too, and every worker pays to spawn and to read the context it shares with the others. This page puts the time, the tokens, and the dollars side by side so you can see the trade before you make it.

Predict first

60 minutes of work, 40% of it serial, 4 workers.

Reading and reconciling each worker’s report takes 3 min. How much faster does it finish than one session doing all of it?

Presets

The first two are the ones where the plausible answer is wrong. Pick one and predict again.

Your scenario

Mode

Subagents are modelled token by token. Teams use the course outline’s multipliers.

Time

Solo duration (minutes)

How long the task takes in one session. An assumption: use your own.

Serial fraction

The share that can’t run in parallel. 40% is the course outline’s example.

Workers

From 1, the solo session, to 32.

Integration per worker (minutes)

Reading and reconciling one worker’s report. An assumption.

Tokens

Token counts take shorthand such as 50k or 1.5M.

Spawn overhead per worker

Tokens before a subagent does anything. Course outline: “somewhere around 7.5k–44k.”

Shared context every worker reads

Files every worker has to read for itself. An assumption.

Unique work input

Split evenly across workers. An assumption.

Output per worker

Report each worker returns

The coordinator reads every one.

Price

Worker model

Dollars per million tokens: input, cached input, output. From the site’s shared price table, editable below.

Cache Workers share a cached prefix

Workers in a fan-out can read the prefix the first one cached, so the other workers pay the cached-input price for their spawn overhead.

Measure spawn overhead from your sessions

Claude Code writes every subagent’s transcript to disk. Choose some, and this page measures what each subagent had in context before it did any work.

Loading the session reader…

When workers spawn workers

A worker with the Agent tool can delegate too. Count what a tree of them adds up to.

Loading the nesting calculator…

Prices

Loading the price table…

How this is calculated

Loading the explanation…