How Much It Costs to Run an AI Agent and How to Measure It

Running an AI agent costs far more than a single model call because the agent loops, and the real driver of the bill is tokens consumed per task, not the price per token. In Anthropic's own data, agents used roughly 4 times more tokens than a chat interaction and multi-agent systems about 15 times more. The number to budget against is cost per successful task, not cost per token, because a cheap model that fails often quietly inflates the true bill through retries.

How much does it cost to run an AI agent?

The cost of running an AI agent is driven mostly by the tokens it consumes, not by the sticker price of the model. An agent does not answer once and stop. It loops through reasoning steps, calls tools, reads what comes back, and reasons again, so it burns far more tokens than a single chat reply for the same task. In Anthropic's own data, agents used about 4 times more tokens than a chat interaction, and multi-agent systems used about 15 times more. That is why the honest unit of cost is not price per token. It is cost per successful task, a figure that folds in retries, failures, and the tokens spent on attempts that never shipped a usable result.

The short version: estimate the tokens one attempt consumes, multiply by how many attempts it takes to succeed, and you have a number you can plan a budget around.

Why token usage, not model price, sets the bill

Founders comparing agents usually start with the price sheet, meaning how many cents per million input and output tokens each model charges. That number matters, but it is the smaller lever. The larger lever is how many tokens the agent spends to finish the job. Output tokens usually cost several times more than input tokens, and an agent that reasons through many steps produces a lot of output, so the loop itself is where the money goes.

Anthropic measured the effect directly. In its analysis of its own multi-agent research system, token usage alone explained about 80 percent of the performance variance on its BrowseComp evaluation, with the number of tool calls and the model choice as the two other explanatory factors. Read carefully, that figure describes one company's internal evaluation of one research task, not a universal law of agent behavior. But the direction it points is hard to argue with. How many tokens you spend swamps almost everything else about how the system performs and what it costs.

The practical consequence is that a cheaper model can be the more expensive choice. If a lower-cost model needs more reasoning steps, more tool calls, or more retries to reach the same answer, it can consume enough extra tokens to erase its per-token discount.

Cost per successful task is the only honest unit

The trap is optimizing for cost per token. A team that does this will happily ship a cheap model that fails, say, two in five tasks, and never notice that the retries and human clean-up cost more than a pricier model that gets it right the first time. That failure rate is a hypothetical for the sake of argument, not a measured figure, but the mechanism behind it is real. Every failed attempt still burns tokens, and someone still has to redo the work.

The fix is to measure cost per successful task, which is roughly the cost of one attempt divided by the success rate. As an illustration, not an industry figure, suppose a task takes 40,000 tokens per attempt and succeeds four times out of five. You pay for five attempts to land four clean results, so your true cost per successful task runs about 25 percent above the naive per-attempt number. The digits here are invented to show the shape of the math. The lesson is that a failure rate you ignore shows up as a cost you pay. This is the same discipline as unit economics anywhere else in a business. If you have never mapped it, our primer on unit economics from day one covers the vocabulary.

A worked example, as an illustration only

Walk one task through, keeping in mind that these numbers are an illustration and not a benchmark for any real workload.

Suppose a single agent run on a support-triage task consumes 40,000 tokens across its reasoning and tool calls, and suppose it succeeds four times out of five, so one run in five has to be retried. To ship 100 clean results you actually run about 125 attempts, which is roughly 5,000,000 tokens rather than the 4,000,000 you would guess if you assumed every run succeeded. That is the 25 percent uplift again, arriving through the back door of the failure rate.

Now change one variable. A model that costs more per token but succeeds nine times out of ten needs far fewer retries. Even at a higher headline price, its cost per successful task can land below the cheaper option. You cannot see that on a price sheet. You can only see it once you measure success rate and token use together.

How to measure and control agent cost

A few levers move the number the most.

Route by difficulty. Send easy steps to a small, cheap model and reserve the frontier model for the hard ones. Most tasks do not need your most capable model at every step. Our model routing and inference cost playbook walks through how to set this up without giving up quality.

Cap the loop. Agents fail expensively when they spin in circles, so put limits on tool calls and reasoning steps and make the agent stop or escalate to a human rather than churn tokens indefinitely.

Cache what repeats. System prompts, tool schemas, and shared context are re-sent on every step, so prompt caching can cut a meaningful share of input-token cost on long-running agents.

Instrument before you optimize. Log tokens and success for every task from day one. You cannot manage cost per successful task if the only thing you measure is cost per token, and most teams discover this the hard way after the first invoice.

What this means for LATAM founders

Avante co-founds AI-native companies for Brazil and LATAM, and this cost math is where a lot of AI products quietly break. Frontier inference is priced in dollars while much of the region earns in reais and pesos, so a model call that looks trivial in a San Francisco pitch deck can be a real margin line in a Brazilian or Colombian P&L. Designing around cost per successful task from the first prototype, rather than discovering it after launch, is what keeps an AI-native product profitable as it scales.

The underlying trend helps. Inference prices have fallen steeply and keep falling, which widens the set of agent use cases that clear their own cost. We track that curve for the region in our LATAM AI infrastructure cost curve. The founders who win are not the ones waiting for models to get cheaper. They are the ones who measure honestly today, then route, cap, and cache their way to a defensible cost per successful task.

Frequently asked questions

How much does it cost to run an AI agent?

It depends mostly on tokens, not on the model's headline price. Estimate the tokens one attempt consumes, then multiply by how many attempts it takes to succeed. Because agents loop through reasoning and tool calls, they use far more tokens than a single chat reply, so the figure worth budgeting is cost per successful task.

Why do AI agents cost more than a single chat completion?

Agents do not answer once and stop. They loop, calling tools and reasoning repeatedly, which multiplies token use. In Anthropic's own data, agents used about 4 times more tokens than a chat interaction and multi-agent systems about 15 times more.

What is cost per successful task?

It is the cost of one attempt divided by the success rate. If attempts fail some of the time, you pay for the failed runs too, so the real cost of each usable result is higher than the raw per-attempt price. It is the only unit that captures retries and failures honestly.

Does a cheaper model always lower agent cost?

No. A model with a lower price per token can still cost more overall if it needs more reasoning steps, more tool calls, or more retries to reach the same answer. Measuring success rate and token use together is the only way to know which model is actually cheaper for the job.