What an AI Agent Actually Costs to Run Per Task

An AI agent’s cost per task has five lines, and model tokens are the smallest of them. Anthropic’s own worked example on its published pricing page, fetched 29 August 2026, puts 10,000 support tickets at roughly 3,700 tokens each on Haiku 4.5 at about $37 in total. That is $0.0037 a ticket. Five minutes of human review on the same ticket costs about $3.61 at US editor wages, which is roughly 975 times the token bill.

Almost every agent business case stops at the API price list, because that is the only line vendors publish per unit. The other four are billed monthly, per seat, per gigabyte, or in salaried minutes, and they decide whether the thing pays for itself. Here is the whole stack, priced from public rates as of August 2026.

The five lines that make up cost per task

A production agent task carries model tokens, tool and API calls, retrieval and storage, observability, and human review. Multiply the first four by your retry rate to get cost per completed task rather than cost per attempt. Human review is usually the largest line by one to three orders of magnitude.

Line	What sets it	Order of magnitude per task
Model tokens	Model tier, context size, reasoning length	Fractions of a cent to a few cents
Tool and API calls	Search, code execution, third-party APIs	Cents
Retrieval and storage	Vector reads, writes, reranking	Fractions of a cent
Observability	Trace volume, seats, retention	Cents, amortised
Human review	Minutes spent, reviewer’s wage	Dollars

Model tokens: the line everyone quotes

Published 2026 rates put the cheap tier at roughly a dollar per million input tokens. Anthropic lists Haiku 4.5 at $1 input and $5 output per million tokens, Sonnet 5 at $2 and $10, Opus 5 at $5 and $25. OpenAI’s pricing page lists gpt-5.6-luna at $0.20 and $1.20. Both vendors discount batch processing by 50%.

Two published mechanics change the arithmetic more than model choice does. Cached input costs 10% of the standard input rate at Anthropic and at OpenAI, so an agent that re-reads the same policy document on every run should be paying a tenth for that context. And Anthropic’s page notes that its 4.7 and later tokenizer produces about 30% more tokens for the same text, which raises cost per task without any headline price moving.

Prices also move on a schedule. Google’s Gemini API pricing, updated 28 August 2026, lists Gemini 3.7 Flash at $0.75 input and $3.75 output per million tokens through 31 December 2026, then $1.50 and $7.50 from 1 January 2027. Your per-task cost doubles on a calendar date you did not set.

The tool bill is separate, and it is not small

Server-side tools are priced on top of tokens. Anthropic’s page lists web search at $10 per 1,000 searches, code execution at $0.05 per container-hour after 1,550 free hours a month, and managed agent sessions at $0.08 per session-hour. An agent that searches three times per task has already spent $0.03 before a single token is billed, which is eight times the token cost of that support ticket.

Tool definitions cost tokens too, on every request. Anthropic publishes the overhead: the computer-use toolset adds about 4,500 input tokens per request and the browser toolset about 6,600. Attach both to a chatty agent and the scaffolding costs more than the work.

Retrieval and observability sit underneath as a floor. Pinecone’s pricing lists storage at $0.33 per GB per month, read units at $16 to $18 per million, and reranking at $2 per 1,000 requests, on a Standard plan with a $50 monthly minimum. LangSmith lists $39 per seat per month on Plus with 10,000 base traces included. Neither publishes a per-trace price, so treat both as fixed costs you amortise across volume.

The line nobody prices: human review

If a person checks the agent’s output, that person is your dominant cost per task. US Bureau of Labor Statistics OEWS data for May 2025, pulled from the BLS public API on 29 August 2026, puts the annual mean wage for editors at $90,060, customer service representatives at $46,590, and management analysts at $113,790. Divide by 2,080 hours and a reviewer minute costs $0.37, $0.72, or $0.91 depending on who reviews.

By my arithmetic that makes five minutes of review $1.87, $3.61, or $4.56 in base wages alone, before benefits, before management overhead, before the cost of the reviewer’s attention being somewhere else. Against $0.0037 in tokens, the token line disappears. Anything that cuts review from five minutes to one is worth more than any model switch you will ever make.

This is why we log every action an agent takes. A reviewer who has to reconstruct what happened spends four minutes doing it and one minute deciding. A reviewer looking at a structured run log spends one minute total. The paper trail is not a compliance feature, it is the cheapest lever on cost per task.

Reliability decides cost per completed task

An agent that succeeds half the time costs you two attempts for every finished job. METR’s Time Horizons benchmark, whose leaderboard was last updated 8 May 2026, measures this directly: it reports the human-expert task length at which a model succeeds 50% of the time, and separately at 80%. For Claude Opus 4.5 those figures are 293 minutes and 49 minutes. The 80% horizon is roughly six times shorter across every model on the board.

Read that as a pricing instruction. Scope each agent to a task well inside its 80% horizon and you pay close to one attempt per completed job. Scope it to the 50% headline and you have doubled your token, tool, retrieval, and review bill at once, because the human reviews the failures too.

Our own fleet of 41 agents ran 5,450 times with zero failures across one measured 13-day stretch, as of July 2026. The reason is narrow scope per agent, not a clever model choice. One of those agents recomputes real contribution margin daily against the ledger, and we published what it found on the ad-spend truth engine page.

Why per-task cost rises while token prices fall

Both things are true at once and the contradiction confuses most estimates. Epoch AI, publishing 12 March 2025, found the price of reaching a given capability level falling between 9 and 900 times per year depending on the task: GPQA Diamond performance went from $37.50 per million tokens in March 2023 to $0.12 by December 2024.

MIT FutureTech’s paper The Price of Progress, revised 23 March 2026, measures the same fall at 5 to 10 times a year for fixed capability, with about 3 times a year attributable to algorithmic efficiency alone. Then it reports the counter-trend: the price of running frontier models is rising 3 to 18 times a year, driven by larger models and longer reasoning traces.

So the cost of last year’s answer collapses while the cost of this year’s answer climbs. If you buy capability you do not need, your per-task cost goes up every quarter. If you fix the task and let the price of that capability fall, it goes down. That choice is worth more than any negotiation with a vendor.

What we do with this

We publish every price and we do not mark up your infrastructure. Your API accounts, your database, your hosting stay in your name, so you see the per-task bill directly rather than through an invoice line. A tomte, our word for one production agent with one defined job, gets scoped narrow enough to sit inside its reliability horizon, logged well enough that review takes a minute, and handed over with the token math written down. An AI Workforce Sprint is $9,500 fixed for that, in four to six weeks.

Before you price the running cost, price the build: we broke that down in what a custom AI agent costs in 2026. And before you price either one, capture the four baseline numbers in how to measure AI ROI, because a cost per task means nothing without the cost of the task today.

Send us the job you want priced and we will write down what it costs to build and what it costs to run, per task, before you commit anything. You get a written scope and a fixed number back within one business day. No meetings. Start async.