What It Actually Costs to Run an AI Agent

Most builders get one of two surprises when they add up what an agent actually costs to run. Either the inference bill is larger than expected, or the tool-call bill dwarfs the inference bill and they planned for neither.

Planning for both up front takes maybe an hour. Discovering them in production takes longer.

The inference layer

Every prompt your agent sends to a language model costs tokens. Every major provider publishes per-token rates, and the arithmetic scales fast with volume. What most builders underestimate is not the per-token price but the context overhead.

An agent that maintains conversation history, stores tool results in context, and loads a multi-tool schema sends far more tokens per turn than a simple prompt-response pair. That same rate applied to a much larger context window produces a much larger bill. Profile your typical context window early; do not estimate it.

Inference costs are the most transparent part of the stack because providers document them. They are also the part of the bill least affected by the market structure described below. Plan for them separately, then add the second layer.

The tool-call layer

Every external call your agent makes is a separate line item. Search, document retrieval, database queries, financial data, code execution, and browser automation each carry their own pricing. Many are per-call.

Across the paid agent endpoints currently tracked in the live index, the median price per call is $0.01. That figure excludes six endpoints with implausible prices and the free endpoints held out of the percentile calculation. It reflects what working, paid services actually charge per request.

A single cent per call sounds small until you multiply it by the number of tool calls per task and the number of tasks per day. For any agent doing more than a handful of tool calls per task at volume, the tool bill rivals or exceeds the inference bill. The tool bill also grows in ways inference does not: each capability you add to an agent tends to add more tool calls per task, not fewer.

Track tool calls per task as a metric from day one, alongside total spend. A workflow whose tool-call count per task is rising is a workflow whose costs will keep rising without a deliberate decision to contain them.

When agents hire other agents

The bill compounds faster when your agent delegates to specialist agents rather than raw API calls. A document-processing workflow might hire a retrieval agent, a structured-extraction agent, and an output-formatting agent for a single job. Each of those has its own per-call price and its own tool calls.

The payment activity is measurable: 430,346 USDC in transfers to agent payment addresses in the trailing seven days, across 818 distinct services. The mandatory caveat: that figure is heavily concentrated. One service accounts for 36.3% of all measured transfers, and the top 10 services are 87.6% of volume. Spending is real but it is not spread evenly across the market.

The concentration has a practical implication for budget planning. Only 14.2% of listed paid endpoints have more than one distinct paying buyer in the last 30 days. Before you design a workflow that delegates to multiple specialist agents, check which endpoints in your category have real buyers behind them and which are listed without a market. A service with no paying customers outside its own operator may not stay maintained once you depend on it.

What a usable budget model looks like

Four rows cover most of what will surprise you:

Inference. Count expected tasks per day. Estimate tokens per full turn, including system prompt, tool schemas, history, and response. Multiply by your model's per-token rate. Add a buffer for retries, error loops, and the tasks that turn into multi-turn conversations.

Direct tool calls. List every external API your agent will call. Get the per-call price for each before committing to a workflow design. Some services have published pricing; others require you to sign up and check. Build a row for each dependency and update it regularly, because pricing changes.

Agent delegation. If your workflow hires specialist agents, get their per-call rates before designing the delegation structure. Treat published rates as floor estimates. Services that find early traction often raise prices once adoption builds; that is worth knowing before your workflow depends on them.

Retry and error overhead. Agents retry. Models fail to follow tool-call formats. Third-party APIs return empty results that trigger follow-up calls. Add a reasonable overhead factor to your total expected costs, and revise it once you have real production data.

The number to track

Cost per successful task, not cost per call.

An agent that makes many tool calls and completes the job is cheaper than one that makes fewer calls, fails to complete the job, and retries multiple times. The failing agent costs more even though its per-call price looks lower.

Completion rate and per-task cost are the same metric viewed from different angles. A workflow with a lower completion rate pays a hidden overhead in retries before you even count the tasks that never resolve.

Track both from your first production run. The ratio between them tells you whether your agent is getting more expensive to operate over time or more efficient.

Where most of the money goes

The payment data, concentrated as it is, suggests that agent-to-agent spending flows mostly to data services: financial feeds, search results, retrieval endpoints. Processing and verification services are smaller categories. Data that an agent cannot generate itself commands pricing power because there is no substitute. Commodity processing work faces more price competition.

Category matters for budget planning. An agent in a data-intensive workflow carries higher tool-call costs than one doing reasoning over data it already has. The ratio shifts depending on what your workflow actually needs, so baseline it before you try to optimize it.

Live pricing for agent endpoints changes as services find their market. The index reflects current crawl data, and any specific number you anchor to today may look different in six months. Check live figures before you finalize a budget for a new workflow, and schedule a cost review for anything running at meaningful volume.