How Much Does an AI Agent Cost? Price the Completed Task

An AI agent costs whatever one completed task costs you, and for most business agents the model's tokens are the smallest part of that number. Price a task the way it actually runs: tokens and tool calls for every attempt, the minutes a person spends checking the output, and the cost of fixing the tasks that go wrong, all divided by the share that finish. This piece is for the owner or operations lead deciding whether an agent pays for itself. It gives you the cost stack, the formula, dated price anchors and a calculator, so you can budget on cost per completed task instead of price per million tokens.

Update, 2 October 2026: Correction: an earlier version said people carry 83% of the unit cost and tokens under 10%. On the calculator defaults, review ($1.50) and failure recovery ($1.00) are $2.50 of the $2.86 spent per attempted task, about 87%, and tokens ($0.288) are about 10%. The chart and the closing section now use these shares; no other figure changed.

Agoda's AI Developer Report 2026, released today, names cost as the top barrier to wider agent adoption, ahead of integration and governance. Anthropic and OpenAI both launched cheaper models two days earlier. Cheaper tokens will not fix a cost problem that is mostly people and rework.

How much does an AI agent cost to run?

An AI agent costs, per task, its tokens and tool calls on every attempt, plus human review, failure recovery and a share of fixed costs. In the Agoda AI Developer Report 2026 (24 September 2026), cost was "the primary barrier to wider adoption, cited by 28% of respondents, ahead of integration complexity (24%) and lack of governance (19%)."

The respondents were developers across seven markets in Southeast Asia and India, not business owners, and the release does not give a sample size. Read the 28% next to two other numbers in the same report. Four in five survey respondents work under "some degree of AI usage limits, such as token quotas, usage caps, or budget restrictions", and 86% still review or validate AI output always or most of the time. So the teams worried about cost are already rationing tokens, and they are still paying people to check the work.

Agoda's chief technology officer, Idan Zalzberg, put it plainly: "The real cost is not simply the model call. It is the work required to make AI trustworthy in production."

The guides that rank for this question mostly quote build fees, monthly ranges and token prices. They rarely price a finished task. Almost none count the reviewer's minutes, which is where budgets go wrong.

A managing director needs the all-in cost of one finished invoice match or one resolved ticket, set against the person doing it today. Everything below is built to produce that number.

What goes into the cost of one completed task?

One completed task carries five cost lines plus one multiplier (retries), and only the token and tool lines reach a provider's invoice. The rest sit in payroll, incident time and other vendors' bills. The table shows what drives each line and the dated price anchor where one exists; the calculator further down has an input for each. Prices are US dollars at list rates.

Retries are not a separate bill. They multiply tokens and tools, and a retry usually costs more than the first attempt because the agent re-reads what it has done so far, plus the error. Anthropic's own Opus 5.5 announcement (22 September 2026) says cache reads, the re-read context, "make up the majority of agentic and coding work costs."

Failure recovery is the line nobody budgets. An agent that sends a wrong quote or files an invoice against the wrong supplier costs far more than the tokens it burned. It costs the twenty minutes a senior person spends finding and fixing it, and sometimes a customer conversation. That cost is real even when nobody records it.

For compliance and risk teams, one point matters more than the arithmetic. Human review shows up in the calculator as a cost, but it is a control, and cutting it to hit a budget is the wrong lever. Reduce the failures that make review necessary, then decide which tasks can be sampled instead of checked one by one. Our piece on whether your human approver is a rubber stamp covers how to keep that check meaningful.

Why does the average cost per task mislead?

The average cost per task misleads because agent costs are uneven: most tasks finish cheaply, while a few eat a large slice of the budget. OpenAI's own usage data, published on 6 September 2026, shows the spread: its median researcher used more than $600 of inference a day by mid-August, while the 90th percentile used more than $7,000.

That is roughly a twelvefold gap between the middle and the top tenth, inside one organisation. It measures spend per person per day, not per task, but the shape carries over. The same report says "agents still require significant human steering to be successful, especially as task complexity rises", and that "over half of successful 4-8 hour tasks involved 1 or more interventions." Harder tasks cost more tokens and more people. Both lines rise together.

Illustrative example: an agent handles 1,000 supplier queries a month. 950 finish in one pass for about $0.30 of tokens each. 45 need a retry and a person to check the answer, about $4 each once review time is counted. Five go wrong in a way someone has to unwind: half an hour of senior time plus tokens burned in a loop, about $45 each.

The median task costs $0.30. The average is $0.69. The five worst tasks, half a percent of volume, take a third of the spend, and the 50 tasks that needed help take nearly 60%.

Budget on the $0.30 and you are wrong by more than double. Budget on the $0.69 and you are right on average, then surprised in the month the five bad tasks become fifteen. That is what a fat tail means in practice. The fix is a ceiling: a spend cap per task, a retry limit, and a rule that any task hitting either goes to a person instead of trying again.

Where to set the ceiling is a judgement call, and no published norm exists. Our starting suggestion, not a standard: cap spend per task at about five times your median and allow two retries at most, then adjust once you see how often the cap fires. A cap that fires daily is too tight. One that never fires is decoration.

For the CTO, this changes what to log. Write one row per task: task ID, attempts, input, output and cache-read tokens, tool calls, review minutes, outcome (passed, fixed or failed) and the minutes any fix took. Then compare the median and the 90th percentile every month. If the gap keeps widening, the tail is growing faster than the volume.

Which AI model prices matter in September 2026?

The prices that matter are the list price of the model you would actually run and the per-call price of every tool it uses. On 22 September 2026, Anthropic priced Claude Opus 5.5 at $4 per million input tokens and $20 per million output, and OpenAI priced GPT-6 Luna at $0.10 and $0.50.

The table sets those beside a mid-tier and a low-cost option, at standard list rates in US dollars per million tokens, from the providers' own pages as they stood on 24 September 2026.

Two cautions apply before you compare rows. A token is not the same size everywhere: Anthropic's pricing page notes that its newer tokenizer "produces approximately 30% more tokens for the same text", so a lower price per token can still mean a higher price per page.

Tools carry their own meter. Web search costs $10 per 1,000 calls at both Anthropic and OpenAI, and every retry pays it again. Batch work is the one easy discount: both companies take 50% off for batch processing, which suits overnight jobs such as reconciliations or report drafting.

The headline is the 40-fold gap between Opus 5.5 and Luna. On a token bill, that gap is decisive. On a completed-task bill, as the next section shows, it can vanish.

How do you calculate AI agent cost per completed task?

AI agent cost per completed task is the total spend per attempted task divided by the share of tasks that complete. The formula behind the calculator is: unit cost = ((token cost + tool cost) x (1 + retry rate) + review minutes / 60 x hourly cost x review share + failure rate x recovery cost) / (1 - failure rate). Monthly cost is unit cost x tasks plus fixed costs.

The defaults are illustrative, not benchmarks. They describe a mid-complexity agent on Claude Opus 5.5 list prices: 40,000 input tokens and 4,000 output tokens per attempt, $0.06 of tool calls, a 20% retry rate, a five-minute check on half of tasks at a loaded $36 an hour, and a 5% failure rate costing $20 each to put right.

With the defaults, the arithmetic runs like this. Tokens cost $0.16 in and $0.08 out per attempt, plus $0.06 of tools: $0.30. A 20% retry rate makes that $0.36. Review adds five minutes at $36 an hour on half of tasks: $1.50. Failures add 5% of $20: $1.00.

That is $2.86 per attempted task, and dividing by the 95% that complete gives about $3.01 per completed task. At 1,000 tasks plus $400 of fixed cost, the month comes to about $3,411. The token line is under a tenth of it.

Illustrative example: keep everything the same, but switch the model to GPT-6 Luna at $0.10 and $0.50. If Luna fails no more often than Opus 5.5 here, the unit cost falls to about $2.71, a saving of 30 cents a task.

Now suppose the cheaper model fails 12% of tasks and averages 1.5 attempts per task, a 50% retry rate. Tokens fall to under a cent per attempt, the model-and-tools line drops to about $0.10, and the unit cost rises to about $4.54. Run the numbers backwards, holding the retry rate at 20%, and the break-even is narrow: Luna wins only while its failure rate stays below about 6.2%, against 5% for Opus 5.5.

That is the decision this calculator exists for. With these defaults, each extra percentage point of failure adds about 24 cents to the unit cost, while the switch to Luna saves 30. So a model 40 times cheaper per token loses if it fails roughly one more task in every hundred. Test the cheap model on your own tasks and measure its failure rate before you switch, because the token price cannot tell you.

For a finance lead, the cheapest lever is often outside the model. Cut the check from five minutes to three by making output easier to verify, say a one-line summary with a link to the source record, and the unit cost falls from $3.01 to about $2.38. That 63-cent saving is double what Luna saves at equal failure rates.

Operators should see their own time in the review line. If the checker spends eight minutes instead of five because the output is hard to verify, the unit cost rises to about $3.96, three times the Luna saving in the other direction. Easy-to-check output is a cost lever.

If you are in Singapore, Budget 2026 added AI expenditure to the Enterprise Innovation Scheme for Years of Assessment 2027 and 2028, capped at S$50,000 a year. This is general information, not financial advice; confirm what qualifies before counting it, as our guide to the 400% AI tax deduction explains.

Where Origin Pi stands

We think an agent is worth running when its cost per completed task, people included, beats today's way of working at the 90th percentile too. In our illustrative defaults, review and failure recovery carry about 87% of the unit cost, so price per million tokens is an input to that test, not the answer.

So we design agents to be measurable per task from day one. Every attempt and correction goes to an audit trail against its task, which replaces the calculator's guesses with your own numbers. Each agent runs with bounded permissions and a spend and retry ceiling, so a task stuck in a loop stops and goes to a person instead of quietly running up a bill. The confirm step sits where a mistake would be expensive to unwind, so review minutes buy down failure cost instead of duplicating it.

That discipline is the core of our business automation work, and why agentic ROI holds up only when governance keeps it honest.

Your next step: pick one task an agent does or could do, and for two weeks record four numbers per task: attempts, review minutes, whether it needed fixing, and what the fix took. Put them in the calculator. If the token line is more than a fifth of the result, optimise the model. If not, optimise the failures.