Token-Based Pricing: How to Account for AI Credits and LLM Usage

Token-based pricing is how most AI companies charge for their services. OpenAI, Anthropic, and similar providers measure usage in tokens, small chunks of text, and bill accordingly. Whether you're buying AI credits or selling AI services, understanding the accounting treatment matters for accurate financial reporting.

Last verified: November 2025

---

What Are AI Tokens

Tokens are the basic unit AI companies use to measure and charge for usage. When you send text to ChatGPT or Claude and get a response, both your input and the AI's output are measured in tokens.

Think of tokens as pieces of words. One token equals roughly:

4 characters of text
3/4 of a word in English
A punctuation mark or space

Real examples:

"Hello, how are you?" = 5 tokens
A 1,000-word article = about 1,300 tokens
A typical chat message = 100-300 tokens

Why tokens matter: AI providers charge based on how many tokens they process. The longer your prompt and the longer the response, the more you pay.

---

How AI Companies Price Tokens

Different models have different prices:

GPT-4 (OpenAI's advanced model):
$30 per million input tokens
$60 per million output tokens

GPT-3.5 Turbo (faster, cheaper):
$0.50 per million input tokens
$1.50 per million output tokens

Claude Sonnet (Anthropic):
Has its own pricing per token

What this means in practice: A 1,000-word article processed through GPT-4 costs about $0.04 for the input and $0.08 for the output (assuming a similar-length response).

---

Common Ways to Pay for AI Tokens

1. Pay-As-You-Go

You get billed monthly for whatever tokens you actually use. No upfront payment.

How it works: Use AI services throughout the month, get an invoice at the end showing your total token usage and cost.

Best for: Companies with predictable, ongoing AI usage

2. Prepaid Credits (Most Common)

You buy credits upfront, then each token you use deducts from your balance.

How it works: OpenAI requires at least $5 in credits to start. Buy $100, use $20 worth of tokens, you have $80 left.

Best for: Getting started quickly, avoiding payment issues

Why companies like it: Reduces failed payments since you've already paid. Automatically cuts off service if you run out (protecting them from unpaid usage).

3. Monthly Subscription with Allowance

You pay a flat monthly fee that includes either unlimited usage (with rate limits) or a set token allowance.

How it works: ChatGPT Plus costs $20/month for access to GPT-4 with generous rate limits.

Best for: Individual users who want predictable costs

4. Auto-Reload Credits

Set up automatic top-ups when your balance drops below a threshold.

How it works: When your credits hit $10, automatically charge $50 to your card and add it to your balance.

Best for: Preventing service interruptions

---

Accounting for Token Credits: If You're the Customer

When You Buy Credits

Prepaid credits are an asset (something you own that has value), not an expense.

Journal entry when buying $1,000 in credits:

Debit: Prepaid AI Credits $1,000
Credit: Cash $1,000

As You Use Tokens

Each time you use AI services, convert the used credits from an asset to an expense.

Journal entry after using $250 worth of tokens:

Debit: AI Services Expense $250
Credit: Prepaid AI Credits $250

Where to categorize the expense: Depends on what you're using AI for:

R&D Expense if building AI products
Cost of Goods Sold if AI powers your customer-facing features
Operating Expense if using for internal tools or content

Month-End Process

At the end of each month:

1. Check how many tokens you used
2. Calculate the cost of those tokens
3. Record the expense even if not invoiced yet

If using pay-as-you-go (billed in arrears):

Debit: AI Services Expense
Credit: Accrued Expenses

Then when the invoice arrives:

Debit: Accrued Expenses
Credit: Accounts Payable

When Credits Expire

Many AI credits have expiration dates (OpenAI's free trial credits expire after 3 months). If yours expire unused, write them off.

Journal entry:

Debit: Other Expenses (or Loss on Expired Credits)
Credit: Prepaid AI Credits

Don't carry expired credits as assets, they have no value once expired.

---

Accounting for Token Credits: If You're the Provider

When You Sell Credits

Prepaid credits are a liability (you owe services), not revenue yet.

Journal entry when customer buys $1,000 in credits:

Debit: Cash $1,000
Credit: Deferred Revenue $1,000

As Customers Use Tokens

Convert deferred revenue to actual revenue as customers consume services.

Journal entry after customer uses $250 worth:

Debit: Deferred Revenue $250
Credit: Revenue $250

Handling Credit Expiration (Breakage)

Some customers buy credits but never use them all before expiration. This is called breakage.

How to account for it: Under ASC 606, estimate what percentage of credits will expire based on historical data. Recognize this breakage revenue proportionally as customers use their credits.

Example: You sell $100,000 in credits. Historical data shows 8% typically expire unused ($8,000 expected breakage).

As customers use 60% of their credits:
Regular service revenue: $60,000
Breakage revenue: $4,800 (60% of the $8,000 expected)
Total revenue: $64,800

When credits actually expire:

Debit: Deferred Revenue (unused amount)
Credit: Breakage Revenue

---

Tax Treatment for AI Token Expenses

For Customers Using Cash Basis Accounting

Can you deduct prepaid credits immediately? Depends on the 12-month rule:

Deduct immediately if: Credits expire within 12 months of purchase or by the end of the next tax year

Must capitalize if: Credits last longer than 12 months past purchase

Example: Buy $5,000 in credits on December 1, 2024 that expire June 30, 2025. The 12-month rule applies, so you can deduct the full $5,000 in 2024.

For Customers Using Accrual Basis Accounting

You must capitalize prepaid credits as an asset. Deduct as expense when you actually use the AI service (economic performance).

Why: Under accrual accounting, expenses are deducted when incurred, not when paid.

Reference: IRS Publication 538

For Providers

Accrual basis: Prepaid credits are not taxable income until you deliver services (as customers consume tokens).

Cash basis: Credits are generally taxable when received, though some deferral may be allowed under specific IRS procedures.

Reference: IRS Publication 535

---

Volume Discounts and Tiered Pricing

Some AI providers offer cheaper rates as you use more tokens in a month.

Example pricing tiers:

First 1 million tokens: $60 per million
Next 4 million tokens: $50 per million
Over 5 million tokens: $40 per million

How to account for this:

For customers: Track which tier you're in to understand your effective cost. If you cross a tier mid-month, your average cost per token changes.

For providers: Your billing system must track cumulative usage and apply the right tier automatically. Revenue recognition follows the actual rates charged.

---

Tracking and Managing Token Usage

For Customers

Set up tracking by:

Project: Which projects are using how many tokens?
Team: How much is marketing spending vs engineering?
Cost center: Allocate costs to the right departments

Use budget controls:

Set spending limits in the AI platform
Configure alerts when approaching limits
Review usage weekly to catch anomalies

Monthly reconciliation:

Compare the provider's usage report to your recorded expenses
Validate token counts and rates
Investigate any discrepancies

Tools like afternoon.co can automate this by importing usage data directly from AI providers and categorizing expenses automatically.

For Providers

Track everything:

Every token processed
Which customer used it
What timestamp
Which model (GPT-4, Claude, etc.)
Input vs output tokens

Maintain accurate records of:

Prepaid credit balances per customer
Usage rates and pricing tiers
Expected breakage based on historical patterns

---

Different AI Services, Different Accounting

LLM API Calls (OpenAI, Anthropic)

Charges apply to both your prompt (input tokens) and the response (output tokens).

Accounting: Expense/recognize revenue as the API call happens

Image Generation (DALL-E, Midjourney)

Usually priced per image, not per token, but bought with the same credit system.

Accounting: Expense/recognize revenue when images are generated

Model Fine-Tuning

Training a custom AI model consumes many more tokens than regular use (often 6x the rate).

Special consideration for customers: If the fine-tuned model will be used long-term, consider capitalizing the training cost as an asset that you amortize over its useful life. Otherwise, expense immediately.

Embeddings and Vector Search

Charges include tokens to process text into embeddings plus ongoing storage fees.

Accounting: Separate treatment for one-time token charges (usage-based) vs monthly storage fees (subscription-based)

---

Common Mistakes to Avoid

Expensing prepaid credits immediately: They're an asset until you use them. Expensing on purchase misstates your financials.

Not tracking expiration dates: Unused credits that expire must be written off, not kept as assets.

Lumping all AI costs together: Track token usage separately from other cloud expenses for better cost analysis.

Missing cost allocation: If multiple teams share one AI account, track and split costs appropriately.

Forgetting month-end accruals: With pay-as-you-go billing, accrue for tokens used but not yet invoiced.

Ignoring breakage (for providers): You must estimate and recognize revenue from credits likely to expire unused.

---

Free Credits and Promotions

Many AI companies offer free credits to new users or through partnerships.

For Customers

Accounting for free credits: Don't record anything when you receive them (you didn't pay). Don't record an expense as you use them (they were free).

What to track: Monitor usage for planning purposes. Once free credits run out, you'll need to understand what paid usage will cost.

For Providers

When you give away free credits:

Option 1: Expense immediately as marketing cost
Option 2: Defer the marketing cost and recognize as credits are used

Journal entry (immediate expense):

Debit: Marketing Expense
Credit: Promotional Credit Liability

No revenue is recognized as free credits are consumed, you're not earning anything from them.

---

Key Metrics to Track

For providers selling token-based services:

Revenue per million tokens
Average monthly tokens per customer
Token usage growth rate
Gross margin on token processing
Breakage rate (% of credits expiring)
Credit utilization rate (% of purchased credits used)

For customers buying token-based services:

Total monthly AI spending
Cost per project or team
Tokens used per outcome (e.g., tokens per article generated)
Average cost per thousand tokens
Percentage of credits used before expiration

---

Summary

---

Official References

• FASB ASC 606 - Revenue from Contracts with Customers
• IRS Publication 538 - Accounting Periods and Methods
• IRS Publication 535 - Business Expenses