Key Takeaways

  • Token-based invoices need separate input, output, and cached-input quantities tied to the model and the rate effective when usage occurred.
  • A discount off a model’s current list price changes with that price; a fixed token rate follows its own contract terms.
  • Prepaid balances need an agreed credit burn order, expiry rules, and a record of which model consumed each amount.
  • An API commitment statement should reconcile opening balance, rated token consumption, adjustments, and remaining spend without charging prepaid usage again.

The customer disputes the cached-input line. Their engineering team counted tokens across its API keys, but the invoice used a new model price for requests made before the price cut. Now Finance has to explain both the token quantity and the rate. Token-based billing for AI API companies puts those questions on the same invoice.

If you sell model access by the token, a monthly total cannot explain the bill. Input and output carry different rates. Cached input has its own price. An enterprise customer can consume several models against the same spend commitment, while a self-serve customer draws from expiring prepaid credits.

For each request, Finance needs the model version, token category, rate in effect, and account paying the bill. Here is how to carry those details through a rate card, handle the invoice exceptions, and calculate a commitment drawdown after a mid-month price change.

How Do AI API Companies Structure Token-Based Billing?

Separate token measurement from how the API customer pays. Monthly arrears, prepaid credits, and committed spend can use the same token measurements. The agreement decides what happens next: which token charges reduce a balance and which go on a new invoice. That changes the amount due.

Pricing component How it is measured What goes wrong in billing
Input and output tokens Separate quantities by model, version, and request. A blended token rate hides a change in the input-to-output mix.
Cached input Cached quantity separated from ordinary input. The same input is charged at both the ordinary and cached rates.
Batch or asynchronous usage Eligible requests with the agreed discounted rate. The batch discount stacks with a contract discount without permission.
Media and fine-tuning Images, seconds, or characters; training tokens and hosting measured separately. Non-token units enter a token price calculation, or hosting disappears.
Reserved capacity and dedicated GPUs Provisioned units per hour or dedicated GPU-hours. Idle reserved capacity is excluded despite the reservation charge.
Prepaid and promotional credits Eligible monetary consumption against each credit balance. The wrong balance burns first or expired credits remain available.
Enterprise commitments and volume tiers Eligible spend across models and the contract’s tier measurement period. List-price spend replaces discounted drawdown, or tiers reset at the wrong boundary.

Where Does the Token Invoice Break?

A price cut overwrites the old rate

Keep model version and rate version separate. A price cut can change the rate without changing the model. Applying today’s token price across the month reprices the earlier requests. A discount off a model’s list price follows changes only as the agreement specifies; a fixed negotiated token rate does not automatically fall with list.

What to check: Retain effective timestamps and the pricing time zone. Split usage at the rate boundary and preserve the list price, discount basis, and net rate used.

The customer counts a different set of tokens

A request total cannot resolve a cached-input dispute. Finance needs the request reference and separate input, cached-input, and output quantities, including whether the reported input total already contains cached tokens. Failed, cancelled, and retried requests also need explicit billing treatment. A retry is not the same thing as a duplicate copy of an event.

What to check: Match the requests the customer disputes to your metered records. Show exclusions, billable retries, and duplicate events removed before applying rates.

Credits expire while usage is still arriving

Paid credits and promotional credits can have different eligible models, burn priorities, and expiry dates. Late usage complicates that order: a request made before expiry may arrive afterward. Using ingestion time without an agreed rule can consume paid funds that should have remained available.

What to check: Define eligibility by event time or another explicit contract rule. Record which credit balance each charge used and how a correction puts it back.

Projects roll up but commitments do not

The paying account can own many projects and API keys. Model-specific charges must reach that account’s eligible commitment without merging unrelated agreements. If cloud marketplace purchases count toward committed spend, reconcile that channel’s credited amount so the same consumption does not create another direct charge.

What to check: Map projects, keys, purchase channels, and eligible products to the agreement. Show each model’s drawdown within the account total.

The term ends with unused spend

A commitment ramp sets when spend is due. Rollover or forfeiture decides what carries into the renewed term. Unused prepaid value and unpaid committed spend need different billing actions. Charging a shortfall already collected as prepayment would collect the same obligation again.

What to check: Compare eligible consumption, amounts already collected, ramp milestones, and overage terms. Calculate any remaining amount due before applying the contract’s rollover or forfeiture rule.

What Does a Mid-Month Price Change Do to the Commitment?

This hypothetical example uses an annual commitment of $12,000, paid upfront. All eligible token usage receives 20% off the list price effective when the request occurs. Earlier consumption was $2,000, leaving a $10,000 opening balance. The example month has 30 days. There are no promotional credits, taxes, other fees, or volume tiers.

Model A’s price changes at the start of day 16 in the contract’s time zone. Its model version stays unchanged. Model B’s prices stay unchanged throughout the month. Quantities below are millions of tokens; each rate is dollars per million tokens. Input means uncached input, so cached tokens are excluded from that quantity.

Model and period Token category Quantity in millions List rate Net rate after discount Charge
A, days 1 to 15 Input 20 $2.00 $1.60 $32.00
A, days 1 to 15 Output 5 $8.00 $6.40 $32.00
A, days 1 to 15 Cached input 10 $0.50 $0.40 $4.00
A, days 16 to 30 Input 30 $1.50 $1.20 $36.00
A, days 16 to 30 Output 10 $6.00 $4.80 $48.00
A, days 16 to 30 Cached input 20 $0.30 $0.24 $4.80
B, days 1 to 30 Input 40 $1.00 $0.80 $32.00
B, days 1 to 30 Output 10 $4.00 $3.20 $32.00
B, days 1 to 30 Cached input 20 $0.20 $0.16 $3.20

Apply the discount: A 20% discount leaves 80% of list, so each net rate equals list rate × 0.80. Before the change, Model A input costs $2.00 × 0.80 = $1.60 per million. Its line is 20 × $1.60 = $32.00. After the change, input costs $1.50 × 0.80 = $1.20 per million, giving 30 × $1.20 = $36.00.

Total the rated lines: Model A before the change is $32.00 + $32.00 + $4.00 = $68.00. Afterward it is $36.00 + $48.00 + $4.80 = $88.80. Model B is $32.00 + $32.00 + $3.20 = $67.20. Total consumption is $68.00 + $88.80 + $67.20 = $224.00.

Show the settlement: The monthly statement shows $224.00 in consumption, $224.00 covered by prepayment, and $0.00 newly due. Remaining balance is $10,000.00 minus $224.00 = $9,776.00. Cumulative consumption is $2,000.00 + $224.00 = $2,224.00; adding the remaining $9,776.00 reconciles to the original $12,000.00 commitment.

The rate split belongs in the customer’s supporting detail even though prepayment covers the month. Repricing earlier Model A requests at the new rates would understate drawdown and overstate the remaining balance.

What Does Finance Need From a Token Billing System?

Require an invoice view grouped by model, token category, and effective rate, with request-level evidence underneath. Reprocessing a duplicate event should leave charges unchanged. Price a late request under the original terms. If you’ve already billed that period, show the adjustment so Finance can trace it.

Keep model-level charges available after commitment drawdown. Model mix matters when serving costs differ. Finance needs to compare the consumption sold for each model with its delivery costs, even when the customer sees a shared credit balance.

For prepaid token access, cash collection and revenue timing can differ. Feed the credit record and the revenue schedule from the same consumption evidence, while reviewing commitments, unused balances, and reserved capacity under the actual agreement. Confirm the treatment with your auditor.

Where Does Vayu Fit for Token Sellers?

Vayu’s usage metering page states: “Vayu helps measure API calls, transactions, tokens, seats, workflows, storage, compute, and other usage events, turning raw data into controlled billable usage.” Its pricing models workflow supports effective dates and customer-specific exceptions, with a preview of the billing impact before a change goes live.

AU10TIX, an identity verification company, combines annual commitments with usage-based pricing. Its case study describes Vayu consumption metering that processes tens of millions of events daily. Read the customer story.

Aquant, an AI platform for field service organizations, bills AI usage through Answers and Minutes Used. Its commercial model combines a fixed AI Platform subscription with prepaid usage and pay-as-you-go consumption. Vayu applies pricing logic across those components. Read the Aquant story.

These stories support metering and mixed commercial structures. Neither is presented as a model API provider. For a token-selling business, test the dated rate card and commitment calculation above against your own agreement.

What Should You Ask a Token Billing Vendor?

  1. Can you reproduce input, output, and cached-input charges on both sides of a model price change?
  2. Can Finance distinguish a fixed negotiated rate from a discount off the effective list price?
  3. Can you replay late and duplicate requests without changing the wrong credit balance?
  4. Can projects and marketplace purchases reconcile to the paying account’s eligible commitment?
  5. Can you show prepaid coverage, overage, and term-end shortfall without collecting the same spend twice?

Frequently Asked Questions

How does token-based billing work?
Separate tokens by model and category, apply the effective contract rates, and settle the charge through credits or arrears. A model API seller needs the rate history as well as the token history.

How do you bill customers for LLM API usage?
Roll requests from projects and API keys into the paying account. Retain input, output, cached quantities, and billable status. For products selling agent outcomes instead, see AI agent billing and financial control.

How do prepaid credits work for AI APIs?
Eligible model charges reduce an available balance. The contract sets credit priority, expiry, and overage behavior. Promotional and paid funds need separate histories so a correction restores the appropriate balance.

What should billing software for AI infrastructure companies handle?
Require dated model rates, token categories, capacity charges, and commitment balances. Reconcile late token usage before close; the month-end close checklist covers the wider close process.

How do you handle a price change during a billing period?
Split requests at the contractual effective timestamp. Apply each period’s list price and discount, or retain the fixed negotiated rate where required. Keep the earlier rate available for disputes and late events.

Can Finance explain the rate behind every token charge? Book a billing logic review and walk through your model rates, credit rules, and commitment statement.