---
title: Why do multi-agent systems use so many more tokens?
url: https://opencontroller.lyzr.com/blog/why-multi-agent-systems-cost-more
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Why do multi-agent systems use so many more tokens?

**Because each agent needs its own context, and coordination itself costs tokens. Anthropic reported that its agents typically used about 4× more tokens than chat interactions, and its multi-agent research system about 15× more. Multi-agent designs pay off on high-value tasks that split into independent parallel parts; for tightly linked work such as most coding, they mostly add cost.**

This post explains where the tokens go, when the cost is justified, and how to measure it. It is part of our [gateways and cost](/blog/llm-gateway-vs-agent-gateway-vs-control-plane) series.

## What did Anthropic actually measure?

In June 2025, Anthropic published an engineering write-up of the multi-agent system behind Claude's Research feature: a lead agent that plans and spawns sub-agents to search in parallel. Key figures it reported:

- Agents typically used about **4×** more tokens than chat interactions.
- Multi-agent systems used about **15×** more tokens than chats.
- On its internal research evaluation, the multi-agent setup (a lead Opus 4 agent with Sonnet 4 sub-agents) outperformed single-agent Claude Opus 4 by **90.2%**.
- On the BrowseComp evaluation, token usage alone explained **80%** of the variance in performance.

The write-up also said multi-agent systems suit tasks where value is high enough to pay for the extra tokens, and that domains where all agents must share the same context, including most coding tasks, are not a good fit today.

## Where do the extra tokens go?

| Cost driver | What happens | Why it multiplies |
|---|---|---|
| Duplicated context | Each sub-agent gets instructions, tool definitions and background | The same tokens are sent N times |
| Tool definitions | Every agent loads its tool schemas | Large MCP tool lists cost tokens on every call |
| Planning and delegation | Lead agent writes task briefs for each sub-agent | Extra generation before real work starts |
| Parallel exploration | Sub-agents search overlapping ground | Some work is done more than once |
| Result passing | Sub-agent findings flow back into the lead's context | Lead context grows with every report |
| Synthesis | Lead agent reads and merges everything | Long-context call at the end |
| Retries and loops | Failed sub-tasks re-run | Errors cost a full sub-agent run |

The [token cost of MCP tool definitions](/blog/token-cost-of-mcp-tool-definitions) alone can be large when several agents each load dozens of tools.

## When is a multi-agent design worth it?

Use this checklist before choosing one:

1. **Is the task parallelisable?** Can it split into parts that do not need each other's intermediate results?
2. **Is the value per task high?** A due-diligence report justifies more spend than a password reset.
3. **Does one context window run out?** If a single agent cannot hold the needed information, splitting helps.
4. **Is latency important?** Parallel sub-agents can finish faster than one agent working sequentially.
5. **Have you tried a single agent with better tools first?** Often cheaper and easier to govern.

If you answer "no" to the first two, a single agent is usually the better choice.

## How do you keep multi-agent costs under control?

1. **Measure cost per task, not per call.** Roll up every model call under one task ID. See [cost per task](/blog/measure-cost-per-task-for-ai-agents).
2. **Attribute spend per agent and per version.** Know whether the lead or the sub-agents dominate cost.
3. **Use smaller models for sub-agents.** Search and extraction often do not need the largest model.
4. **Trim tool lists per sub-agent.** Give each one only the tools it needs.
5. **Compress sub-agent results.** Ask for short structured findings, not full transcripts.
6. **Cap fan-out.** Set a maximum number of sub-agents and calls per task.
7. **Cache shared context.** Provider prompt caching can cut the cost of repeated instructions.
8. **Compare against a single-agent baseline.** Run both on the same evaluation set and compare quality per dollar.

## How Opencontroller does this

Opencontroller's LLM Gateway attributes cost per agent and per version, and supports routing, fallback and retry, and caching. Its October 2026 benchmark measured 23.5% lower model cost with caching on mixed synthetic traffic, under stated test conditions. The registry tracks agents, tools and MCP servers, so teams can see which sub-agents load which tools. See the [LLM gateway benchmark](/llm-gateway-benchmark) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is a 15× token increase typical for every multi-agent system?**
No. It is the figure Anthropic reported for its research system. Your multiple depends on fan-out, context size, models and tool lists. Measure your own.

**Do multi-agent systems always give better answers?**
No. Anthropic reported large gains on research tasks that parallelise well, but noted poor fit for tasks needing shared context, such as most coding.

**What is the quickest way to cut multi-agent cost?**
Use smaller models for sub-agents, give each sub-agent only the tools it needs, and cap the number of sub-agents per task.

## Key facts

Anthropic reported (June 2025) that agents typically used about 4x more tokens than chat, and its multi-agent research system about 15x more; a lead Opus 4 with Sonnet 4 sub-agents outperformed single-agent Opus 4 by 90.2% on its internal research eval, and token usage explained 80% of BrowseComp variance. Extra tokens come from duplicated context, tool definitions, delegation, overlapping exploration and synthesis. Multi-agent suits high-value, parallelisable tasks; most coding is a poor fit.

## Related

- [How do you measure cost per task for an AI agent?](https://opencontroller.lyzr.com/blog/measure-cost-per-task-for-ai-agents.md)
- [How much do MCP tool definitions add to token costs?](https://opencontroller.lyzr.com/blog/token-cost-of-mcp-tool-definitions.md)
- [How do you attribute LLM costs to individual AI agents?](https://opencontroller.lyzr.com/blog/how-to-attribute-llm-costs-to-ai-agents.md)
- [More on Gateways and cost](https://opencontroller.lyzr.com/blog/cluster/gateways-and-cost)
