BroadRouter
Get started

What multi-agent setups really cost: per token vs flat sessions

By the BroadNet Inference team · 4 min read · Published · Updated

A multi-agent setup costs roughly the tokens of every agent added together, plus handoffs and repeated context. Anthropic measured about 15 times the tokens of a chat for its multi-agent research system, against 4 times for one agent. Per token, the bill grows with every agent; on a flat plan it grows with the agents running at once.

Splitting a task across an orchestrator and sub-agents can finish faster and explore more. It also multiplies the bill, in ways that aren’t obvious from a price list. This post works out what a multi-agent setup costs per token and on flat sessions, using BroadNet’s published prices for DeepSeek V4.1 Flash. How to count and buy sessions is covered in running parallel coding agents on a flat rate; this post is about the money.

How many more tokens does a multi-agent system use?

The clearest public numbers come from Anthropic’s write-up of its multi-agent research system. It reported that “agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats.” It also found token usage alone explained 80% of the variance in its performance tests. More agents spend more tokens, and often get better results for it.

The same write-up is careful about fit: “For economic viability, multi-agent systems require tasks where the value of the task is high enough” to pay for the extra performance. It also notes most coding tasks have fewer truly parallel parts than research.

So a reasonable planning assumption is that a multi-agent version of a task costs 3 to 4 times the single-agent version. Your design can be cheaper or much worse. Measure it.

Where do the extra tokens come from?

Multiplier Why it happens
Per-agent overhead Each sub-agent sends its own system prompt and tool definitions on every call
Duplicate reading Sub-agents reread the same files because they don’t share context
Result passing The orchestrator reads every sub-agent’s output as new input
Separate caches Different prefixes per agent mean fewer cached input hits
Retries A failed sub-task resends its whole context

The agent token budget post shows how to measure each bucket per call.

How do the two billing models scale?

The key difference is what each one counts:

  • Per token (pay as you go) counts every token every agent sends and receives. Add an agent, add its tokens.
  • Flat sessions count requests in flight at the same time. On BroadNet’s unlimited plans, one session runs one request at a time with no token meter, for $60 a month on Full Moon. Four agents calling the model at once need four sessions. Four agents taking turns need one.

That means the plan doesn’t make multi-agent free. A session only produces so much: at about 100 tokens per second, about about 259 million output tokens a month if it generated non-stop. More tokens in the same hours means more sessions.

The session math, worked

We use the full-time coding agent from our flat-rate vs per-token post: 1,000M input tokens a month, 80% cached, 50M output. On DeepSeek V4.1 Flash that costs $65.40 a month per token.

Setup Monthly tokens Per-token bill Sessions Plan bill
One agent, all day 1× $65.40 1 $60
Orchestrator + 3 sub-agents, all busy all day 4× 4 × $65.40 4 4 × $60
Same team, sub-agents in bursts, peak overlap 2 about 2× 2 × $65.40 2 2 × $60
Four light agents, side-project volume each 4 × light 4 × $14.23 4 4 × $60

What the rows show:

  1. Busy agents favor the plan, by a margin per session. Each fully loaded session saves $5.40 a month at this volume. Four sessions save four times that.
  2. Bursty sub-agents favor fewer sessions. Buy for the overlap you see most of the time. BroadNet’s pages don’t say whether a request above it waits or is refused, so give every agent a retry with backoff for a concurrency 429.
  3. Light agents favor per token. Four sessions for four mostly idle agents cost far more than their tokens. At side-project volume, each agent costs $14.23 a month per token.

The rule of thumb: a session is worth it when the agents using it push about 917 million input and 46 million output tokens a month through it.

How do you estimate your own bill?

  1. Run the multi-agent workflow for a week on a pay-as-you-go key, with usage logging per agent.
  2. Sum tokens per agent: input, cached input and output. Multiply by about 4.3 for a month.
  3. From request timestamps, find how many requests overlap most of the time. That’s your session count.
  4. Compare the per-token total against sessions × $60 in the flat-rate vs per-token calculator.
  5. Before buying a month, run the same workflow on $1 Meteor hours, one per session.

A pay-as-you-go key allows 4 requests in flight and 30 a minute, so a large fan-out will also hit rate limits on per-token billing. Plan for backoff either way; see long-running agents.

How do you cut multi-agent cost?

  • Start with one agent. Add sub-agents only where work splits cleanly.
  • Give sub-agents fewer tools and shorter prompts than the orchestrator.
  • Return summaries, not transcripts, from sub-agents to the orchestrator.
  • Use a cheaper model for sub-agents where quality allows, and keep the stronger one for planning.
  • Run handoffs in sequence when speed doesn’t matter. Sequential agents share one session.

Where BroadNet Inference fits

BroadNet’s unlimited plans fit multi-agent systems whose agents stay busy: each session is $60 a month with no token meter on DeepSeek V4.1 Flash, and one key can hold several sessions. Per-token keys fit bursty fan-out and other models such as GLM 5.3. The plan doesn’t fit agents that idle most of the day, and BroadNet doesn’t list tool calling for DeepSeek V4.1 Flash, so test your sub-agents’ tool use first. The OpenAI Agents SDK guide shows the wiring, and plans are at inference.broadnet.ai/plans.

Questions

How much more do multi-agent systems cost than a single agent?

It depends on the design, but Anthropic's published numbers give a scale. Its agents used about 4 times the tokens of chat and its multi-agent research system about 15 times, so roughly 3 to 4 times a single agent for the same question.

Is a flat-rate plan cheaper for multi-agent systems?

When each session stays busy. A BroadNet Full Moon session pays off at about 917 million input and 46 million output tokens a month on DeepSeek V4.1 Flash. Sessions that sit idle most of the day cost more than paying per token.

How many plan sessions does an orchestrator with sub-agents need?

One per request running at the same time. An orchestrator that waits while three sub-agents work in parallel needs three or four, depending on whether it calls the model during that time. Handoffs that run one after another need one.

Why do sub-agents use so many tokens?

Each sub-agent has its own context. It loads its own system prompt and tool definitions, rereads files other agents already read, and returns results the orchestrator then reads again.

Are multi-agent systems worth it for coding?

Sometimes. Anthropic's write-up notes that most coding tasks involve fewer truly parallel parts than research, and that tasks needing shared context fit poorly. Try a single agent first and add sub-agents where work really splits.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity