Cost per Task, Not per Token: AI Unit Economics
Every AI price list is written per million tokens. Nobody in the business buys tokens. They buy a contract reviewed, a ticket answered, a report drafted. A price per token is an input to the cost of that work, in the same way a price per kilowatt-hour is an input to the cost of running a factory. It is not the cost.
Tokens, Explained for the People Who Pay the AI Bill covers what a token is and why the same request can cost very different amounts. This post turns the token into a unit that finance, product owners and SAM teams can manage: cost per task, per user and per feature.
The metrics the FinOps Foundation recommends
The FinOps Foundation's FinOps for AI overview (updated 17 February 2026, accessed 23 October 2026) lists the measures its working group uses for AI. Among them are cost per inference, token consumption metrics including cost per token, cost per API call, return on investment or value for AI initiatives, and time to achieve business value. It also proposes a measure of how well the model chosen matches the quality the task needs.
Cost per token and cost per call describe the supply side. They tell you whether the price you pay is falling. Cost per completed task describes the demand side. It tells you whether a feature is worth running. You need both, and only the second one can be compared with the value of the work.
From list price to landed cost
The list price assumes every request succeeds once, at the standard rate, with nothing cached. Real workloads differ in four ways:
- Caching lowers it. Repeated instructions and reference documents read from a cache at a fraction of the input price, with a premium on the write. Four Vendors, Four Meters: AI Pricing Compared lists each vendor's cache rates.
- Batch lowers it. Anthropic, OpenAI and Google take 50% off for work that can wait hours rather than seconds.
- Retries raise it. A request that fails validation, times out or returns an unusable answer is billed, and so is the retry. Peter Walker of OpenRouter, writing in OnlyCFO, gives the arithmetic: a $1 request that succeeds 80% of the time costs $1.25 per success.
- Failed requests can be billed in full. OpenAI's reasoning guide warns that a request can reach its output limit before any visible answer, after paying for the input and the reasoning. Google's Vertex AI price list says that from 13 November 2026 a Gemini 4 completion longer than 32,000 tokens is charged for the tokens generated even if the request then fails on a timeout. The rule is covered in Gemini's API Price Doubles on December 1. Check Your Exposure.
Walker calls the effective price after caching, waiting and retries the landed cost. It is the number to put in a business case.
A worked example
An illustration, not a benchmark or a forecast. The task is a first-pass contract review on Claude Sonnet 5.5, using Anthropic's list prices accessed 23 October 2026: $2 per million input tokens, $10 per million output tokens, cache reads at 0.05 times input ($0.10 per million) and five-minute cache writes at 1.25 times input ($2.50 per million). The assumptions:
- Each review sends a 20,000-token block of instructions and playbook that is the same every time, plus a 30,000-token contract, and gets back 2,000 output tokens including reasoning.
- The team completes 10,000 reviews a month, and the cache is written 500 times a month.
- One attempt in ten fails validation and is retried, billed in full.
| Step | Arithmetic | Per review | Per month |
|---|---|---|---|
| 1. List price, no caching | 50,000 × $2/M + 2,000 × $10/M = $0.100 + $0.020 | $0.120 | $1,200 |
| 2. Shared block read from cache | 20,000 × $0.10/M + 30,000 × $2/M + 2,000 × $10/M = $0.002 + $0.060 + $0.020 | $0.082 | $820 |
| 3. Add cache writes | 500 × 20,000 × $2.50/M = $25 a month | $0.0845 | $845 |
| 4. Landed cost with 10% failed attempts | $0.0845 ÷ 0.9 | $0.0939 | $939 |
Caching took about 32% off the list price. Cache writes and retries together gave back about a third of that saving. If the reviews can wait for an overnight run, the Batch API halves the token rates; step 1 at batch prices is $0.060 per review. Check with the vendor how batch and caching combine for your workload before counting both.
Now divide by people. If 25 lawyers use the service, $939 a month is about $37.56 per user, nearly twice the $20 monthly price of a Claude Enterprise seat. That is the seat-versus-meter gap that Microsoft's AI Bill Has Three Meters describes, measured on one feature.
Model routing
Not every step needs the strongest model. Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. A triage step that reads the 30,000-token contract and writes 300 tokens costs:
- On Haiku 5.5: 30,000 × $0.10/M + 300 × $0.50/M = $0.00315.
- On Sonnet 5.5: 30,000 × $2/M + 300 × $10/M = $0.063.
If triage finds that four contracts in ten are on standard terms and need no full review, the month becomes 10,000 triage steps at $0.00315 ($31.50) plus 6,000 full reviews at $0.0939 ($563.40): about $595 instead of $939. Same illustrative assumptions, same list prices.
Routing has costs of its own. Two models must be tested and monitored, and a triage step that sends the wrong contracts down the cheap path costs more in rework than it saves. Re-measure tokens per task when you change models, too: Anthropic states that the tokenizer in Claude 4.7 and later produces about 30% more tokens for the same text, so a lower price per token does not guarantee a lower price per task. Prices move in both directions as well. Gemini's API Price Doubles on December 1 is the current counter-example.
The monthly report finance wants
A unit cost is only useful if it arrives every month in a shape finance can use. Five views cover most questions:
| View | What it shows | Why it matters |
|---|---|---|
| By team | Seats assigned and meter spend, side by side | Showback, and who to call about a spike |
| By model and tier | Spend and tokens per model, standard, batch or priority | Routing and tier choices, price changes |
| By use case or feature | Completed tasks, landed cost per task, trend | Whether the feature is worth running |
| Seat versus meter | Fixed seat cost, metered cost, cost per active user | Whether the seat or the meter is the real price |
| Committed versus on demand | Prepaid pools and credit packs against consumption | Waste on one side, overage on the other |
Each line needs the unit, the count and the cost, and a comparison with the previous month. Anomalies get an owner and a note. The FinOps Foundation recommends showing this to teams before charging it back to them.
The question this leaves open
Every view in that report needs the same thing underneath it: seats, meters and keys tied to a team, a model and a use case. Most organizations have the invoice and the console exports, but not the join between them. Who owns that join, and where does it live? If You Cannot Read the AI Bill, You Have No Position argues that without it there is no negotiating position either.
To see your own AI spend per person, Software Inventory Manager (SIM) now connects to Anthropic, OpenAI, Google Gemini, Microsoft 365 Copilot and GitHub Copilot: read how on the AI cost management page. To start from the price lists, read Four Vendors, Four Meters: AI Pricing Compared.
