Adding AI features to your SaaS? Here's how to choose between prepaid credits, usage-based billing and a hybrid plan, with real cost maths and what Stripe Billing Meters and Metronome support in 2026.
If you're adding AI features to a SaaS product, start with a hybrid plan: a monthly subscription with an included AI allowance, plus metered overage above it. Move to prepaid credits only when you have several AI actions with very different costs, or when you need cash up front. On Stripe, basic Billing Meters can run the hybrid model today, but real-time credit burndown (stopping a customer the moment their balance hits zero) needs Metronome, the usage-billing platform Stripe acquired.
That's the short answer. The rest of this guide explains why, with real numbers, and shares what I've learned owning Stripe subscription billing, including usage-based metering, at an Australian SaaS company.
What Is the Difference Between AI Credits and Usage-Based Billing?
All three models charge for AI, but they feel very different to the customer.
Usage-based billing (pay as you go): the customer uses the product, and you bill them afterwards for what they consumed. The unit might be AI requests, generated documents, minutes of audio or tokens.
Prepaid credits: the customer buys a balance up front. Each AI action draws credits from it, and different actions can cost different amounts of credits.
Hybrid: the customer pays a fixed monthly plan that includes an AI allowance (for example, 500 generations). Anything above the allowance is metered and billed as overage.
|
Why Do AI Features Break Flat SaaS Pricing?
In classic SaaS, serving one more user costs you almost nothing. That's why a flat $29 a month plan works. AI changes this, because every request calls a model, and every call has a real cost.
Here's a simple worked example. Groq lists Llama 3.3 70B at around $0.59 per million input tokens and $0.79 per million output tokens (as of October 2026; always check
current pricing). Say one AI action uses 2,000 input tokens and 1,000 output tokens:
Input: 2,000 tokens costs about $0.0012
Output: 1,000 tokens costs about $0.0008
Total: roughly $0.002 per action
That looks tiny until you multiply it. A customer who runs 10,000 actions a month costs you about $20 in model fees alone. On a $29 plan, after payment processing and hosting, one heavy user can wipe out the margin for that account. Ten heavy users can wipe out the margin for the month.
I saw this cost shape first-hand while building PrepAI, an AI exam-prep platform I'm developing with Groq and Llama. Generating a test is one LLM call. Grading each answer is another. A single study session can trigger a dozen calls, and a keen student might run several sessions a day. The cost doesn't come from how many users you have. It comes from how much each user does.
So the real question isn't "should we charge for AI?" It's "how do we connect what we charge to what each customer actually costs us, without scaring them away?"
When Should You Choose AI Credits?
Credits work well when:
You have several AI actions with very different costs (a short summary versus a 20-page report, or a cheap model versus an expensive one)
You want cash up front, which helps early-stage cash flow
You want to give away credits for trials, referrals or promotions
Your buyers are teams that like a fixed budget they can top up
Credits are spreading quickly. PricingSaaS counted 79 companies using credit-based pricing at the end of 2025, up from 35 a year earlier.
But credits have a real downside: they hide the price. When a customer sees "this report costs 40 credits", they have to do maths to know what they're paying, and many assume they're being overcharged. Growth advisor Elena Verna summed up the frustration in a
widely shared post titled "I hate AI credits pricing".
If you choose credits, reduce the confusion:
Publish what each action costs in credits, in plain numbers
Show the remaining balance inside the product, not just on the billing page
Send alerts at 50, 80 and 100 percent of the balance
Let customers top up in one click
When Is Plain Usage-Based Billing Better?
Pure pay-as-you-go is the most honest model when your product has one clear unit of value and your buyers are technical. Think of an API that charges per request, or a transcription tool that charges per minute. The customer understands exactly what they pay for, and you never lose money on heavy users.
The risk is unpredictability. Founders and finance teams hate a bill they can't forecast, and a surprise invoice is one of the fastest ways to lose a customer. If you go usage-based, give customers a live usage dashboard, spending alerts and an optional monthly cap.
Why Is Hybrid Pricing the Safest Default for Most SaaS?
A hybrid plan gives you the best of both. The customer gets a predictable monthly price, and you get protection against the small group of users who would otherwise destroy your margin.
A typical setup looks like this:
Starter: $29 a month, includes 200 AI generations
Pro: $99 a month, includes 2,000 AI generations
Above the allowance: a clear, published price per extra 100 generations
Most customers stay inside their allowance and never think about usage. Heavy users pay for what they use. You can size each allowance from your real cost per action, so every plan stays profitable.
Tiers do a lot of the work here. When I built the Starter, Pro and Fleet subscription tiers for Wheelyst, the tiers separated individual users from business accounts running whole fleets. AI allowances work the same way: the tier sets the expectation, and the meter catches the outliers.
What Does Stripe Actually Support in 2026?
This is where many founders get caught out, because Stripe's usage-based billing has changed a lot in the last two years.
Billing Meters (basic usage-based billing)
Billing Meters are Stripe's standard way to record usage. You create a meter (it can sum, count or keep the last value), attach it to a metered price, and send meter events from your backend whenever a customer uses something. Stripe adds it all up and bills it on the invoice.
Stripe also supports billing credits (credit grants) on metered prices. But there's an important catch in the docs: credits are only reconciled at invoice time, so customers can go past their balance during the billing cycle. If you need a hard stop at zero, basic Meters won't give you that on their own.
If you're starting from scratch, my guide to setting up Stripe subscription billing (https://osamahabib.com/blog/stripe-subscription-billing-nextjs-15) covers the subscription side that metering builds on.
Metronome
Stripe completed its acquisition of Metronome in January 2026, and its docs now point new usage-based integrations to Metronome. It handles what basic Meters can't: real-time credit burndown, prepaid drawdown, committed spend and ramps, and different rates per dimension (for example, one price for a small model and another for a large one).
It isn't a drop-in replacement for everything yet. Checkout support is limited, and some Stripe features, including Connect, aren't supported. If your product depends on those, check the current compatibility list before you commit.
Legacy usage records are gone
If your billing still uses Stripe's old usage records API, it's on borrowed time. The legacy API was removed in Stripe API version 2025-03-31. basil, so any integration that upgrades its API version has to move to Billing Meters.
A quick decision guide:
Pay as you go, or subscription plus overage, on a single product: Billing Meters
Prepaid credits that must stop at zero, enterprise commitments, or per-model rates: Metronome
Heavy reliance on Checkout or Connect: start with Billing Meters and review Metronome later
What Owning Usage-Based Billing Taught Me
At Emergi Mentors, an Australian mentorship SaaS, I own the Stripe billing end to end: plan management, Checkout sessions, cancellations, usage-based metering and webhooks. I also built the FastAPI modules for Usage Tracking, Purchase History and Subscription Management. A few lessons apply whichever pricing model you choose:
Record usage in your own database first. Stripe is where the bill is made, but your database should be the source of truth for what each customer did. If you run a multi-tenant product, track usage per tenant from day one (https://osamahabib.com/blog/multi-tenant-saas-nextjs-prisma-postgresql).
Make every usage event and webhook idempotent. Networks fail and Stripe retries events. Without idempotency, retries turn into double charges.
Reconcile before invoices finalise. Compare your own usage totals with what Stripe is about to bill. Catching a mismatch before the invoice goes out is far cheaper than a refund and an apology.
Show customers their usage. Most billing complaints are really surprise complaints. A usage bar in the product prevents most of them.
None of this is glamorous, but it's the difference between billing that quietly works and billing that generates support tickets every month.
Which Pricing Model Should You Pick? A 5-Question Checklist
Answer these five questions honestly and the right model usually becomes clear:
How many AI actions do you have, and how different are their costs? One action points to usage or hybrid. Many actions with very different costs point to credits.
Who is the buyer? Developers accept pay-as-you-go. Small businesses and non-technical teams prefer a predictable plan.
Do you need cash up front? If yes, prepaid credits help.
Is a hard stop at zero acceptable to your customers? If you need one, plan for Metronome or build the limit in your own backend.
Do you rely on Stripe Checkout or Connect? If yes, Billing Meters are the safer choice today.
For most early-stage SaaS, the answer lands on hybrid: a clear plan, a generous allowance, and fair overage.
Frequently Asked Questions
Are AI credits better than usage-based billing?
Neither is better in every case. Credits suit products with several AI actions at different costs and customers who prefer prepaid budgets. Usage-based billing suits a single, clear metric. Most early-stage SaaS should start with a hybrid plan: a subscription with an included allowance plus metered overage.
Does Stripe support prepaid AI credits?
Yes. Stripe billing credits can be granted to customers and applied to metered prices. With basic Billing Meters, credits are reconciled at invoice time, so a customer can use more than their balance during the billing cycle. Real-time burndown needs Metronome.
Should I use Stripe Billing Meters or Metronome?
Stripe now points new usage-based integrations to Metronome, which handles real-time credit burndown, commitments and per-model rates. Billing Meters still fit simple pay-as-you-go or subscription-plus-overage pricing, and they work with Checkout and Connect, which Metronome does not fully support yet.
Can I still use Stripe usage records?
Not on Stripe API version 2025-03-31.basil or later, where the legacy usage records API was removed. Older integrations need to migrate to Billing Meters.
How do I stop AI costs from eating my margin?
Work out your cost per AI action, price above it with a clear margin target, include a capped allowance in each plan, meter usage above the allowance, and send customers usage alerts before they reach their limit.
Building AI Features and Stripe Billing Into Your SaaS?
If you're a founder or CTO adding AI features and usage-based billing to your product, and you're hiring a full stack engineer who has already shipped this in production, I'd like to hear about the role. You can see the billing work I've done (https://osamahabib.com/services/subscription-platform), download my CV (https://osamahabib.com/docs/Osama_Habib_CV.pdf), or get in touch here (https://osamahabib.com/contact).
Osama Habib
Multan, Pakistan
Full Stack Developer specialising in Next.js, Node.js, and the MERN stack. I write about modern web development, system design, and practical engineering.
