What One AI Workflow Actually Costs: Unit Economics

Vendor pricing pages quote dollars per million tokens. No operations budget is denominated in tokens. The question a finance team actually asks is what it costs to process one invoice, resolve one ticket, or review one contract.

This converts current published rates into that unit.

---

Method

Every figure below is arithmetic on list prices published by Anthropic and OpenAI, read in August 2026 and linked in the sources. Token counts per task are stated assumptions, not measurements — substitute your own and the arithmetic holds.

Cost for any task is input tokens times the input rate, plus output tokens times the output rate. That is the whole model.

---

Current rates used

Model	Input per 1M	Output per 1M
Claude Haiku 4.5	$1.00	$5.00
Claude Sonnet 5	$2.00	$10.00
Claude Opus 5	$5.00	$25.00
gpt-4o-mini	$0.15	$0.60
gpt-4o	$2.50	$10.00

Anthropic discounts batch processing 50 percent on both input and output, and prices prompt cache reads at 0.1x base input.

---

Cost per completed task

Workflow	Assumed tokens	Model	Cost per task	Batch (50%)
Support ticket conversation	~3,700 total	Haiku 4.5	0.37 cents	0.19 cents
Invoice extraction, 1 page	1,800 in / 300 out	Haiku 4.5	0.33 cents	0.17 cents
Invoice extraction, 1 page	1,800 in / 300 out	Sonnet 5	0.66 cents	0.33 cents
Contract review, 30k-token doc	30,000 in / 2,000 out	Sonnet 5	8.00 cents	4.00 cents
Contract review, 30k-token doc	30,000 in / 2,000 out	Opus 5	20.00 cents	10.00 cents

The support-ticket figure is Anthropic's own worked example: roughly 3,700 tokens per conversation on Haiku 4.5, stated as about $37 per 10,000 tickets. The others are computed from the rate table above using the stated token assumptions.

---

What the numbers actually say

Document length sets the order of magnitude. A contract review costs 24 times an invoice extraction on the same model. That gap is entirely input tokens. Before comparing vendors, ask whether the task genuinely needs the whole document in context — chunking, pre-filtering or routing only the relevant sections is a larger saving than any model switch.

Model choice is a 2–3x lever, not a 20x one. Within a workflow, moving between Haiku 4.5 and Sonnet 5 doubles cost; moving to Opus 5 triples it again. Real, but second-order against document length.

Batch halves everything that is not interactive. Invoice processing, contract review and overnight classification are all batchable. A workflow that never waits on a human should not be paying interactive rates.

At these unit costs, volume matters less than people expect. One hundred thousand invoices a month on Haiku 4.5 is $330, or $165 batched. The build and integration cost dominates the inference cost at almost any realistic volume.

---

Building your own unit cost

Count the tokens for one real example of the task — not an estimate, an actual run with the usage figures the API returns. Multiply by the published rates. Then multiply by monthly volume.

Two adjustments matter after that. If a stable system prompt or document preamble repeats across calls, prompt caching drops that portion to a tenth of input price after the first call. If the work is asynchronous, apply the batch discount.

What this model deliberately excludes: retries, failed extractions needing a second pass, human review time, and the engineering cost of building the pipeline. In most deployments the human review line is larger than the inference line, which is the strongest argument for measuring accuracy before optimising cost.

---

Working out whether an AI workflow pays for itself at your volumes, or building the pipeline that runs it? Ortem Technologies' AI and ML solutions practice works on both. See our LLM cost optimization guide for cutting an existing bill, or talk to our team →.