A production multi-agent system typically costs $0.50 to $3.00 per agent run for a moderately complex task on a frontier model — and 90-98% of that is model tokens, not platform fees. Open-source frameworks (LangGraph, CrewAI, AG2) cost nothing to license but carry roughly 0.25-0.5 FTE of platform engineering once you run them in production; managed services (Amazon Bedrock AgentCore, Azure AI Foundry Agent Service, Vertex AI Agent Engine, LangGraph Platform) add only about 1-5% on top of your token spend. The real decision is almost never license cost — it's whether your engineers' time is cheaper than the managed premium at your volume.

This guide assumes you have already decided to build with multi-agent AI. If you are still deciding whether the architecture is justified at all, read our analysis of when multi-agent systems are worth the cost first.

What does a multi-agent AI platform actually cost per agent run?

Cost per agent run = (input tokens x input price) + (output tokens x output price) + orchestration compute + platform fees. For a realistic orchestrator-plus-three-subagents research task on a mid-tier frontier model priced around $3 per million input tokens and $15 per million output tokens:

The two lines that matter are input and output tokens. Everything the platform vendor charges is rounding error by comparison — which is why vendor tier tables are close to useless as a budgeting tool.

This is also why multi-agent is structurally expensive. Anthropic's engineering team reported that "agents typically use about 4x more tokens than chat interactions, and multi-agent systems use about 15x more tokens than chats," and that token usage alone explained 80% of performance variance in their evaluation (Anthropic Engineering). You buy capability with tokens, and your architecture decides how many you burn.

How do you calculate cost per agent at your real volume?

Model it in four layers, in this order:

1. Token cost per run. Instrument a real run in staging and log actual input/output tokens per agent. Context grows on every turn, so estimates typically undercount input tokens by 2-4x.
2. Run volume per month. Multiply by successful runs plus retries and failed loops. Budget 15-30% retry overhead in year one.
3. Orchestration and platform fees. Compute-hour rates (AgentCore Runtime is $0.0895 per vCPU-hour and $0.00945 per GB-hour, per AWS pricing), per-node-execution fees, memory/state storage, and tool-call charges.
4. Ops labor. The line vendors never quote and self-hosters always forget.

At 50,000 runs per month with the example above: roughly $70,500 in model tokens, about $1,300 in node-execution fees, and a few hundred dollars of standby compute. Now route the same workload to a small model at $0.15/$0.60 per million tokens and the token bill drops under $5,000 — a 14x swing driven entirely by model choice, not framework choice. Model routing is the highest-leverage cost decision in a multi-agent system; platform selection is a distant second.

OSS frameworks vs managed multi-agent platforms: how do they compare?

Read this as two clusters. OSS frameworks trade cash for engineering time and preserve portability. Managed platforms trade a small percentage premium for eliminated ops work — and buy that with real switching cost, because it is the surrounding services (memory, gateway, identity, tool connectors) that lock you in, not the agent logic itself.

Should you self-host an OSS framework or buy a managed platform?

Run the break-even. A 2% managed premium on $70,500 of monthly token spend is about $1,450. Self-hosting the same workload realistically consumes 0.25-0.5 FTE of a platform engineer for state persistence, retry semantics, queue backpressure, tracing, and on-call — call it $2,500-$5,000 per month fully loaded.

Self-hosting only wins below roughly $1M in annual token spend if you already have a platform team with spare capacity. Below that line, you are paying senior engineering salary to rebuild checkpointing a vendor charges a tenth of a cent for.

Three situations flip the answer toward OSS self-hosting regardless of the math:

• Data residency or air-gapped requirements that managed control planes cannot satisfy.
• Non-standard execution semantics — human-in-the-loop workflows measured in days, or custom scheduling that managed runtimes bill as standby time.
• Multi-cloud or model-agnostic mandates where single-vendor tool lock-in is a board-level risk.

One flips it toward managed: if you cannot name the engineer who will be on-call for agent state corruption at 2am, you do not have the ops capacity to self-host.

What hidden costs blow up multi-agent AI budgets?

Four line items account for most overruns we see in cost reviews:

• Runaway loops. An agent that re-plans instead of failing can burn 10x its normal token budget in one run. Hard-cap turns and total tokens at the orchestrator level, not the agent level.
• Context accumulation. Input tokens grow quadratically in long conversations. Prompt caching and aggressive context pruning typically cut input cost 50-80% and are the cheapest optimization available.
• Observability spend. Tracing every agent step in a hosted platform can rival compute cost. Sample non-critical traces.
• Evaluation and human review. Multi-agent output needs more verification, and reviewer time rarely appears in the platform business case.

The consequences are documented: Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls (Gartner, June 2025). Cost modeling before commitment separates the 40% from the rest.

Which multi-agent platform should you choose?

Apply these criteria in priority order:

1. Where does your data already live? If you are deep in one cloud, the native agent service removes an entire integration surface — worth more than a 2% premium.
2. How deterministic is your workflow? Explicit graph control flow is cheaper to debug and cheaper to run than emergent role-based delegation, because you can see where tokens go.
3. What is your spend trajectory? Under $10K/month, optimize for developer velocity and pick managed. Over $100K/month, invest in routing, caching, and possibly self-hosting.
4. What is your exit cost? Keep agent logic in portable framework code and treat the managed platform as runtime only.

Start managed, instrument aggressively, and revisit at six months with real per-run token data. Teams that self-host on day one usually optimize the wrong layer.

FAQ

How much does it cost to run one AI agent per month? For a single agent handling 1,000 moderately complex runs per month on a frontier model, expect roughly $300-$1,500 in tokens plus under $50 in platform fees. Swapping to a small model for routine subtasks commonly cuts that by 80-90%. Volume and model choice drive the number far more than which platform you use.

Is LangGraph free? The LangGraph framework is open source and free under an MIT license, with no usage limits when self-hosted. LangGraph Platform, the managed runtime, is billed per node execution plus deployment standby time, and higher tiers require a paid LangSmith seat. You can build on the free framework and add the managed runtime later without rewriting your graphs.

Do managed agent platforms charge on top of model tokens? Yes, but usually a small amount. Azure AI Foundry Agent Service charges nothing for the agent service itself — you pay only for model tokens and any tools or knowledge connectors invoked. AWS Bedrock AgentCore bills separately for runtime compute, gateway invocations, and memory operations, which together typically land under a few percent of token spend.

Are multi-agent systems more expensive than single-agent ones? Substantially. Anthropic measured multi-agent systems using about 15x more tokens than chat interactions, versus about 4x for single agents. The architecture only pays back on high-value, parallelizable tasks where the outcome is worth the token multiple.

What is the biggest hidden cost in multi-agent AI? Unbounded retry and re-planning loops. A single misconfigured agent can consume more tokens in one run than a hundred normal runs. Enforce hard turn and token caps at the orchestrator, and alert on per-run cost outliers from day one.

When does self-hosting an open-source agent framework become cheaper? Roughly when annual token spend passes $1M and you already employ a platform team that can absorb the ops work. Below that, the managed premium is almost always cheaper than the fraction of an engineer's salary consumed by state persistence, retries, and on-call.