how to price AI SaaS credit packs

To price AI SaaS credit packs, calculate the fully loaded cost of each billable action, convert those actions into understandable credit amounts, and set pack prices that cover usage, payment fees, and your target margin. Test the model against expensive requests, retries, and heavy users before offering volume discounts or subscriptions.

Key takeaways

• Price completed customer actions, not just raw model tokens. Include infrastructure, retries, tools, storage, and support allowances.
• Use credits to simplify billing, but show customers what common tasks cost before they buy.
• Calculate margins assuming every purchased credit gets used. Unused balances should not rescue an unprofitable product.
• Offer a few straightforward packs, with discounts limited by your worst-case cost per credit.
• Maintain a server-side credit ledger, reserve credits before expensive jobs, and define refunds and expiration clearly.

How do you calculate the real cost of an AI SaaS action?

Start with one outcome a customer recognizes: a generated image, a document extraction, or a completed research report. “One AI request” is usually too vague because two requests can have very different costs.

Build a cost sheet before setting prices

1. Measure model usage. Record input tokens, output tokens, model selection, and cached-input usage where applicable.
2. Add external tools. Include search APIs, OCR, transcription, browser sessions, and third-party data.
3. Add execution costs. Allocate compute, database writes, file storage, and bandwidth.
4. Include unsuccessful work. Measure retries, timeouts, moderation calls, and jobs that consume resources but produce no usable result.
5. Budget operational overhead. Include a reasonable allowance for refunds and usage-related support.

Use provider usage logs from OpenAI, Anthropic, or Google alongside application traces in Langfuse or OpenTelemetry. Check current provider pricing rather than copying token rates from an old tutorial.

As an illustrative internal estimate, an action might cost $0.018 in model usage, $0.007 in tools, and $0.005 in infrastructure and retry allowance: $0.03 total. Those are planning assumptions, not published vendor prices. The common pitfall is treating the $0.018 model charge as the entire cost.

How should you convert tokens and tasks into customer-facing credits?

Use tokens internally and task-based credits externally when possible. Customers can budget for “five credits per document” more easily than a variable bill based on hidden prompts and output length.

Define a small action catalog

• Standard action: One short generation with a defined input and output limit.
• Advanced action: A higher-quality model, longer document, or additional processing.
• Agent workflow: A bounded sequence of searches, tool calls, and generation steps.

For example, you could charge one credit for a short rewrite, five for a document summary, and twenty for a bounded research workflow. These numbers only work if measured costs support them; they are not universal conversion rates.

If that five-credit summary costs $0.03 to deliver, its delivery cost is $0.006 per credit. Repeat the calculation across your catalog. A twenty-credit workflow costing $0.30 consumes $0.015 per credit and may undermine pricing built around summaries.

Keep the action catalog in a versioned configuration table in PostgreSQL or Supabase. Display the credit charge beside each action, and show an estimate for variable jobs. Avoid charging an identical amount for a two-page PDF and a 500-page PDF unless strict limits make their costs comparable.

What formula should you use to price AI SaaS credit packs?

A useful pack-pricing formula includes both percentage payment fees and fixed transaction fees. First choose a target contribution margin: the share of revenue remaining after the variable costs included in your calculation, before fixed business expenses.

Pack price = (credits × delivery cost per credit + fixed payment fee) ÷ (1 − target contribution margin − percentage payment fee).

Assume a pack contains 1,000 credits, each costing at most $0.006 to deliver under your defined workload limits. Using an illustrative payment fee of 2.9% plus $0.30 and a 70% target contribution margin, the calculation is:

($6.00 + $0.30) ÷ (1 − 0.70 − 0.029) = approximately $23.25.

A $25 selling price leaves about $17.98 after the $6 delivery cost and $1.03 payment fee, or roughly 71.9%. This is not net profit: engineering, legal expenses, acquisition costs, and other fixed costs remain.

Build the formula in Google Sheets and test several consumption patterns. Verify your actual Stripe or Paddle fees, including international payments, currency conversion, and billing-product charges. Never use an average cost of $0.006 if customers can freely select actions that cost $0.015 per credit.

Should you sell prepaid packs, subscriptions, or pay-as-you-go credits?

Choose based on how predictably customers use the product. Prepaid packs suit occasional work; subscriptions suit recurring workflows; metered billing suits customers comfortable with variable invoices.

These offers illustrate structures, not validated market prices. Test willingness to pay through customer interviews and actual purchases.

Stripe Billing supports subscriptions and usage-based billing. Paddle provides merchant-of-record capabilities for eligible products, but confirm its current terms and supported billing model. Compare total fees rather than payment processing alone.

Start with one primary model and one top-up option. Launching subscriptions, unlimited plans, rollover, and custom enterprise credits simultaneously makes both customer communication and financial reconciliation harder.

How do you set pack sizes, discounts, and expiration rules?

Offer two or three pack sizes tied to recognizable workloads. Explain both the unit price and the approximate number of standard actions each pack buys.

1. Choose an entry pack. It should support a useful outcome without requiring a large commitment.
2. Set a routine-use pack. Size it around observed monthly consumption, not an arbitrary large balance.
3. Limit the bulk discount. Recalculate the margin at the discounted price with full redemption.
4. Publish balance rules. Explain expiration, rollover, refunds, and whether credits are transferable.

Using the earlier $0.006 delivery cost, a $10 pack of 350 credits leaves about 70.1% after delivery and the illustrative 2.9% plus $0.30 payment fee. A $25 pack of 1,000 leaves about 71.9%; a $90 pack of 4,000 leaves about 70.1%. Larger packs do not automatically produce better margins once discounts enter the equation.

Use Stripe Checkout for purchases and show the same terms inside your billing dashboard. Do not assume purchased credits can expire however you choose: gift-certificate, consumer-protection, and unclaimed-property rules may matter depending on jurisdiction and product structure. Get legal advice before setting expiration policies.

How do you stop agents, retries, and failed jobs from destroying margins?

For agent products, a credit price is incomplete without execution limits. A workflow that keeps searching or retrying indefinitely can cost far more than its advertised charge.

Enforce budgets before starting work

• Set maximum input size, output tokens, tool calls, runtime, and retry count.
• Reserve the estimated credit charge before dispatching a job.
• Settle the final charge when the job completes and release unused reservations.
• Require approval before exceeding the displayed budget.
• Apply per-account concurrency limits and alerts for unusual consumption.

For example, a twenty-credit research job might have a $0.12 delivery budget under the earlier cost model. If realistic testing shows it needs $0.30, charge more credits, narrow the workflow, or change the underlying tools. Do not quietly absorb an unbounded difference.

Use PostgreSQL transactions for balance updates, idempotency keys for charges, and a queue such as Amazon SQS for execution. Redis can help enforce rate limits, but should not be the only durable billing record.

Distinguish infrastructure failures from completed work the customer dislikes. Publish when credits are automatically restored. Test duplicate webhooks, worker crashes, and partial completion: a queue retry must not become a second customer charge.

How do you launch, measure, and adjust credit pricing?

Launch with an explicit hypothesis about costs, value, and usage. Then review actual contribution margin by action type and customer cohort—not just total pack sales.

1. Instrument every job. Record action version, credits charged, estimated provider cost, completion status, and refunds.
2. Review costly outliers. Examine long documents, repeated retries, and heavy tool usage separately.
3. Watch customer friction. Track checkout abandonment, failed balance checks, repeat purchases, and billing questions.
4. Change one pricing variable at a time. Separate pack-price experiments from action-credit changes.
5. Communicate revisions. Give notice and explain how existing balances will be treated.

PostHog can track purchase behavior; Metabase can report margins from your database. Budget analytics and observability subscriptions as overhead rather than assuming free tiers will always suffice.

For a US launch, review state sales-tax obligations and whether your payment setup handles calculation, collection, and remittance. Credit purchases, usage, and recognized revenue may differ in your books; ask a CPA about treatment. If you operate as a sole proprietor, IRS self-employment tax generally concerns net self-employment earnings, not each credit transaction.

If affiliates promote your packs, require clear FTC-compliant disclosures of their paid relationship. Neither affiliates nor your pricing page should promise customer earnings; results vary with effort and market conditions.

Frequently asked questions

What profit margin should AI SaaS credit packs target?

There is no universal target; the 70% contribution margin above is a planning example, not an industry benchmark. Choose a margin that covers your remaining expenses while keeping the product competitive and valuable.

Should one credit equal one token?

Usually not for customer-facing applications, because token costs vary by model and by input versus output. Task-based credits are easier to understand, while developer tools may benefit from transparent token-based billing.

Should unused AI credits roll over?

Rollover can suit uneven usage, but it creates outstanding service obligations and potentially concentrated future demand. Model full redemption, publish any caps, and distinguish purchased top-ups from monthly included credits.

Can you change how many credits an action costs?

Yes, subject to your customer agreements and applicable law, but changing rates reduces the purchasing power of existing balances. Provide advance notice and consider preserving prior rates for previously purchased credits.

Is a free credit allowance better than a free trial?

A limited credit allowance works when customers can experience the core outcome quickly. Set a dollar budget—for example, $0.30 of delivery cost per verified account—and use account and rate limits to control abuse.

Muhammad Tariq at Tariq Services builds AI SaaS products, agents, and automations, including usage tracking and credit-billing workflows. Explore AI development services if you need help connecting pricing rules to reliable product infrastructure.