Inference Cost Calculator
Cost live model traffic from your request volume, token counts and prices.
In short
Formula: C = n x (tin / 1,000,000) x Rin + n x (tout / 1,000,000) x Rout + Platform
What this calculator does
Metered quantity times rate, with the token counts averaged over the period by the user. Input and output are costed separately. Cost per thousand requests follows the FinOps unit-economics definition.
Use it to turn Currency, Requests in the period, Input tokens per request, Output tokens per request, and the other shown inputs into a checked result you can compare, copy, or rerun with different assumptions.
The page shows the formula, a numeric worked example, and the assumptions that affect this professional & industry calculation.
Inputs and what they mean
- Currency
- — Amounts are shown in the currency you pick. No exchange rate is applied..
- Requests in the period
- — How many model calls were made..
- Input tokens per request
- — Average input tokens in one call..
- Output tokens per request
- — Average output tokens in one call..
- Price per million input tokens
- — Take this from your own bill or the provider's price page. No provider price is built in..
- Price per million output tokens (optional)
- — Optional component. Leave at 0 to exclude it from the result. Take the rate from your own bill or the provider's price page; no provider price is built in..
- Platform or gateway cost (optional)
- — Optional component. Leave at 0 to exclude it from the result..
- Price as of (optional)
- — The date you took these prices. It is shown with the result so the figure is never read as a current vendor price..
- Price source (optional)
- — Where the prices came from. Shown with the result..
How to use it
- Enter your own figures — the calculator never fills in a rate, price or benchmark for you.
- Press Calculate to see the result.
- Read the formula, variables, assumptions and source below the result before you rely on it.
Formula
C = n x (tin / 1,000,000) x Rin + n x (tout / 1,000,000) x Rout + Platform.
Inputs used: Currency, Requests in the period, Input tokens per request, Output tokens per request, Price per million input tokens, Price per million output tokens (optional), Platform or gateway cost (optional), Price as of (optional), and other shown inputs.
Edge handling: Requests in the period is required.; Requests in the period cannot be negative.; Requests in the period must be no more than 1000000000000.; Input tokens per request is required.; Input tokens per request cannot be negative..
Worked example
250,000 calls of 1,200 in and 350 out
- Start with Currency: USD, Requests in the period: 250000, Input tokens per request: 1200, Output tokens per request: 350, Price per million input tokens: 2.50, Price per million output tokens (optional): 10.
- Apply the method: C = n x (tin / 1,000,000) x Rin + n x (tout / 1,000,000) x Rout + Platform.
- Token counts measured from your own logs, prices from your own bill.
Reading the result
The headline figure is the main answer. Any breakdown underneath shows the parts that make it up, so you can check the working and see what changes when you adjust an input.
Limitations and assumptions
- Inputs outside the supported range are rejected rather than forced into a result.
- The result depends on the values you enter for this inference cost calculator; it does not supply missing rates, rules, prices, dates, or assumptions for you.
Reference: The Linux Foundation — Unit economics — FinOps Framework
Last reviewed:
Common questions
Where does this formula come from?
FinOps Framework capability: unit economics — cost per business unit is total cost for a period divided by the units delivered in the same period; a workload's cost is the sum of its metered components charged at the rates the organisation actually pays. See the source link on this page.
What kind of calculation is this?
Request volume multiplied by the tokens each call uses and the per-million-token prices, with any platform charge added. Check the formula, example, and limitations on this page before using the result for a real professional & industry decision.
What are its limits?
Arithmetic over the figures entered — it cannot tell whether the metering, logs or invoices behind them are complete. The result is only as current as the rates entered; a rate taken months ago is not a current price. Tiered, committed-use, cached-input and minimum-spend pricing are not modelled unless the user has already reduced them to the rates entered. An average token count hides the spread; a long tail of large requests can move the real bill. Retries and failed calls are only included if the user counted them as requests.
How do I use the Inference Cost Calculator?
Enter the required values for Currency, Requests in the period, Input tokens per request, Output tokens per request, Price per million input tokens, and the other fields shown. The calculator applies the formula on this page and shows the main result with any supporting breakdown so you can check the arithmetic.
What formula does the Inference Cost Calculator use?
C = n x (tin / 1,000,000) x Rin + n x (tout / 1,000,000) x Rout + Platform The visible formula section above lists the calculation path and the edge cases the page handles, so the result can be checked without relying on the form alone. Check the formula, example, and limitations on this page before using the result for a real professional & industry decision.
Can the Inference Cost Calculator be used for exact decisions?
Use it as a calculation aid, not as a substitute for checking the underlying rule, contract, policy, or professional advice that applies to your situation. When a result depends on local rules, personal details, prices, or dates, enter those values yourself and confirm them before acting.
Formula, source and verification
Request volume multiplied by the tokens each call uses and the per-million-token prices, with any platform charge added.
The question it answers: What are we paying for the tokens our application sends and receives?
The formula
C = n x (tin / 1,000,000) x Rin + n x (tout / 1,000,000) x Rout + Platform
- n — Requests
- (requests). Model calls served in the period.
- tin — Input tokens per request
- (tokens). Average input tokens in one call.
- tout — Output tokens per request
- (tokens). Average output tokens in one call.
- Rin — Price per million input tokens
- (currency). From the user's own bill or price page.
- Rout — Price per million output tokens
- (currency). Optional; excluded at zero.
- P — Platform or gateway cost
- (currency). Optional; excluded at zero.
Units: Requests and tokens in, currency out, with a currency-per-thousand-requests figure.
What kind of calculation this is
Business input model. The answer depends on business figures only you can supply, such as your own costs, rates or volumes. No market or benchmark values are assumed for you.
Method
Metered quantity times rate, with the token counts averaged over the period by the user. Input and output are costed separately. Cost per thousand requests follows the FinOps unit-economics definition.
Assumptions built into the result
- Mathematical: All the figures entered cover the same measurement period.
- Mathematical: Token counts, requests and hours are taken from the user's own logs or bill.
- Mathematical: The average token counts entered represent the traffic in the period.
Figures this calculator will never guess for you
- No target, benchmark, industry average or typical figure is supplied for any value.
- A definition is never presented as a goal: a cost per request is not a target cost per request, and a contribution margin is not a target margin.
- No model price, GPU price, rate card or list price is built in; every amount is entered by the user.
- No token count, context length or output length is assumed on the user's behalf.
- No currency conversion is applied; amounts stay in the currency selected.
Limitations
- Arithmetic over the figures entered — it cannot tell whether the metering, logs or invoices behind them are complete.
- The result is only as current as the rates entered; a rate taken months ago is not a current price.
- Tiered, committed-use, cached-input and minimum-spend pricing are not modelled unless the user has already reduced them to the rates entered.
- An average token count hides the spread; a long tail of large requests can move the real bill.
- Retries and failed calls are only included if the user counted them as requests.
Source and version
- Standard or reference
- FinOps Framework — unit economics and workload cost — FinOps Foundation (FinOps Framework capability: unit economics — cost per business unit is total cost for a period divided by the units delivered in the same period; a workload's cost is the sum of its metered components charged at the rates the organisation actually pays.)
- Published source
- The Linux Foundation — FinOps Framework
- Formula version
- Version 1
- Verification
- Reviewed against the cited source on
- How much weight the source carries
- Standards or government
- Applies to
- Currency
- Amounts stay in the currency you choose; no exchange rate is applied.
Related tools
Work out the hourly rate to charge a client so that target margin is met after overhead and non-billable time.
Find the true hourly cost of an employee once benefits, payroll cost and overhead are included.
Size a software application from its data and transaction functions before estimating effort or cost.
Estimate development effort, schedule and average team size for a software project of a known size.
Express cloud spend as a cost per customer, per transaction or per other business unit.
Set the rate used to apply manufacturing overhead to jobs for the coming period.