The Unit Economics of an AI Agent

The AI-agent economy has one number it quotes obsessively and another it almost never computes. The first is volume: x402, an agent-payment rail that barely existed a year earlier, passed 165 million cumulative transactions by late April 2026 and has kept climbing through the summer, the large majority of them on Base. The second is the income statement of a single agent doing a single task—what one job actually costs to run and what it can be sold for. Every deck describes the rails. Almost none builds the P&L.

The received wisdom, which we have argued ourselves, is that most agent tokens behave like memecoins because the agent has no cash flow for the token to capture. That is correct, and it is also unfinished: it asserts the missing cash flow rather than computing it. This article builds the statement line by line—the cost of a task, the multiplier everyone leaves out, the margin that survives, who pays for the inference, and why cheaper models do not save you. The token question answers itself once the arithmetic is on the table.

(Prices below are mid-2026 API rates. Model names and numbers drift monthly; treat them as a dated snapshot, not a constant.)

The unit, and the multiplier everyone ignores

The unit is one task: one agent completing one job—a research report, a resolved bug, a booked trip, an executed trade. The instinct is to price it as one model call. That instinct is wrong by an order of magnitude, and the error is the whole story.

An agent task is not a call. It is a loop: plan, call a tool, read the result, critique it, retry, call another tool, assemble an answer. Anthropic, describing its own multi-agent research system, put hard numbers on this—a single agent uses about 4× the tokens of a chat interaction, and a multi-agent system about 15×, with token usage alone explaining roughly 80% of the performance variance on their research evaluation. More loops buy more quality, and they cost proportionally.

Worse than proportionally, in fact. Each loop re-sends the growing transcript: on step twenty the agent re-reads the previous nineteen steps before choosing the twentieth. Cost scales like the sum of a growing series—closer to quadratic in steps than linear—so each successive call carries more input than the last. That inflates input volume quickly, though whether it dominates the bill depends on the price ratio: with output priced ~5× input and a decent cache-hit rate, generated tokens often still cost more than re-sent ones. Both climb with the loop, which is the point.

The cost of one task

Start with the call. In mid-2026 a frontier model runs about $5 per million input tokens and $25 per million output (Claude Opus 4.8; GPT-5.6’s flagship tier sits slightly above at $5/$30; Gemini’s Pro tier undercuts at roughly $2/$12). A cheap frontier-lab tier is a fifth of that—$1 in, $5 out (Claude Haiku 4.5). Open-weight models served on a US host land lower still, around $0.50–1.70 per million (DeepSeek V3, Llama 3.3). A single frontier call of five thousand input and one thousand output tokens costs about five cents.

Now run the loop, and check it against measured reality rather than the five-cent call. A deep-research task—the kind that reads dozens of sources and writes a report—costs about $0.95 in inference, measured on a mid-tier model (the ~$2/$12 band) with roughly a 60% cache-hit rate. Run the identical task on the frontier tier quoted above and it lands closer to $2.10. A coding agent runs about $1.05 per benchmark instance attempted, and nearer $1.69 per bug actually resolved—a distinction worth keeping, since failed attempts still burn inference. The loop turned a nickel into a dollar or two. Note that is a wider spread than the 4–15× band above: the chat-turn baseline is a smaller call than the five-cent one priced here, and a research task sits at the deep end of the loop. Note too that the tier you pick moves the total by more than any prompt optimization will. Caching helps, but only on the input side, so on a task whose bill is mostly generated tokens it shifts the total by tens of percent, not multiples. That is before the task’s other line items.

Three things fall out immediately. The rails everyone writes about—x402, the micropayment, the wallet—are a rounding error; settlement is sub-cent against a dollar of inference. On a search-heavy task the tooling can cost more than the model—grounded search at retail rates passes a dollar for eighty queries—which inverts the cost structure most decks assume. And the cost of goods is not a number, it is a distribution with a fat tail: the median task clears at a dollar, and the occasional task that spirals into a long loop costs ten. Anything that prices the task at a flat rate is short the tail. And a task that fails outright still burns its full inference while earning nothing—rework is a cost line the median hides, and under per-outcome pricing it compounds with the token tail.

The revenue side, and the margin that survives

Against that cost, what can an agent charge? The public reference points are low. Salesforce prices agentic work two ways: about $2 per agent conversation, and $0.10 per action—the latter being the discrete unit of work proper. A research task is closer to a conversation than an action, so $2 is the fairer comparison, and it is still low. Put a $2 price against a $1.05–2.15 median COGS and the gross margin on a hard task runs from roughly 48% down through zero—before the tail is counted. A cheap, tool-light task clears comfortably; a search-heavy one on a frontier model loses money at that price. For the tasks actually worth automating, the margin is thin, volatile, and sometimes absent.

And it gets thinner, because the task is usually commoditizable. If the agent’s only input is a model any competitor can also call, there is no moat, and price falls toward cost plus a sliver. The durable margin lives in the layers a competitor cannot copy: proprietary data, a hard-won tool integration, distribution to the buyer, and the willingness to carry liability for the outcome. An agent that is a thin wrapper on a public model has the unit economics of a reseller marking up a commodity, which is to say almost none.

Why cheaper models do not save you

The obvious objection: inference is collapsing in price, so the margin problem solves itself. For an LLM of fixed quality, cost has fallen roughly 10× a year—a16z’s “LLMflation”—and Epoch AI measures the drop for a fixed capability at anywhere from 9× to 900× a year depending on the task. Surely the agent’s COGS follows it down.

It does not, and the reason is the multiplier. As models get cheaper they also get more capable, and more capable agents are handed longer horizons—more steps, deeper loops, always-on operation. The direction is structural rather than a proven time series, but the incentive is one-way: capability gets spent on longer tasks, so per-task token draw climbs even as the per-token price falls. Forbes made the arithmetic concrete in July 2026: if a task’s token draw rises 20-fold while the unit price falls 75%, total model charges still rise fivefold. (Market-wide, Goldman projects token consumption growing 24× by 2030—that figure is adoption, not per-task intensity, but it points the same way.) LLMflation is real, and it is out-run by loop-growth before it reaches the agent’s income statement. The price of a token is not your cost; the price of a token times k is, and k is climbing.

Who pays for the inference

There is a principal-agent problem baked into every agent’s P&L, and it is decided by one question: who eats the variable, fat-tailed COGS?

If the user pays per call—the inference passed through—the agent has no incentive to bound its own loop. Every extra retry is revenue for the model provider and cost for the user, and “thoroughness” becomes a euphemism for an unbounded bill. If the agent charges a fixed price and eats the inference, it carries the 30× tail risk itself, and the temptation is to protect margin by clipping loops—fewer steps, worse answers, quality quietly cut to defend a spread. If it charges per outcome—paid only for a resolved task—incentives on quality align, but the agent is fully exposed to the expensive tail, and one pathological task can erase the margin on ten clean ones.

Subscription and capped-usage plans—a fixed base with metered overage—are the common real-world hybrid, but they only relocate the problem: at the cap, the clip-the-loop incentive comes back. There is no free option. The billing model is not a pricing detail bolted on at the end; it is the business model, and it is downstream of the COGS distribution. Pick it before you build, because it determines which tasks you can afford to take.

Where the token captures value

Now the token question answers itself. The agent’s revenue and its settlement are dollars—x402 clears in USDC, Google’s AP2 is deliberately payment-agnostic (cards and bank rails included, stablecoins merely among them), and the inference bill is denominated in dollars too. The margin is thin and volatile. A project token layered on top of this flow captures none of it, because none of the unit economics runs through the token. That is the arithmetic under the memecoin diagnosis: not a narrative failure, a cash-flow one.

A token earns its place only when it is load-bearing inside the unit economics, and there are only a few honest positions:

Everything else—revenue-share promises, governance over an agent that earns nothing, a “unit of account” for volume that does not exist—fails the same test: run the task P&L, and check whether a single dollar of margin actually passes through the token. Usually it does not. The four value-capture mechanics and the pattern catalog are the menu; this P&L is the test each dish has to pass before it is worth cooking. On the other side of the same dollar, the inference cost is a decentralized-compute operator’s revenue—the supply-side economics we cover in the DePIN piece.

A unit-economics template for an agent

This is the cost-of-goods layer that sits beneath the usual LTV and CAC—the line most agent models skip, because measuring it turns a clean story into a distribution with a tail.

The takeaway

An AI agent is priced one task at a time, and the task is a loop, not a call—so its cost is the per-call price times a multiplier that runs from four to fifteen and spikes to thirty, denominated in dollars, with a fat tail that flat pricing ignores. Against that, the sellable price is low and the margin thin wherever the work is commoditizable, and falling model prices do not rescue it because capability spends the savings on longer loops. The token captures value only where it is load-bearing in that statement—as bonded reliability, mandated settlement, or paid access—and nowhere else. The x402 transaction counter will keep climbing. The number that decides whether any of it is a business is the one almost no one prints: the margin on one agent, doing one task, on an average Tuesday.