Why are multi-agent pipelines so expensive?

Each hop between agents is a fresh LLM call that re-includes context, and tool calls add schema plus result tokens on top. A task that touches five agents over three hops can fire a dozen billable calls.

A hop is one handoff between agents in the workflow — one round of an agent thinking, possibly calling a tool, and passing output to the next agent. More hops means more billable turns.

How do I account for reflection loops?

Reflection loops repeat agent turns to critique and improve output. Model them by raising the average hops per task, since each reflection adds another round of calls.

Is my data sent anywhere?

No. The calculator runs entirely in your browser. Nothing you enter is uploaded, stored or logged.

Multi-Agent Pipeline Cost Estimator

Multi-agent pipeline cost estimator

Multi-agent frameworks like AutoGen, CrewAI, and LangGraph make it easy to chain many LLM calls together — and just as easy to lose track of the bill. Every hop between agents is another billable call, every tool invocation adds overhead tokens, and reflection loops multiply it all. This estimator turns your pipeline shape into a concrete cost per request and per day.

How it works

The model treats a task as a series of hops. On each hop an agent produces a turn worth of tokens and may call a tool, which adds schema and result tokens. The total tokens per task scale with the number of hops and the per-turn size:

tokens/task = hops × (avg_tokens_per_agent_turn + tool_overhead)
cost/task   = tokens/task / 1,000,000 × price_per_1M
cost/day    = cost/task × daily_tasks

The agent count drives how rich each turn tends to be and helps you sanity-check the hop estimate — a five-agent DAG rarely resolves in a single hop.

Tips and notes

Cap the hops. Set a hard limit on handoffs and reflection rounds; runaway loops are the number-one cause of multi-agent cost blowups.
Use cheaper models for routing agents. Reserve the expensive model for the agent doing the real reasoning; routers and validators can run on a mini model.
Trim tool schemas. Verbose tool definitions are re-sent on every hop — shrinking them compounds across the whole pipeline.