The model that made SaaS great is the model breaking AI SaaS
Flat subscription pricing won for a reason. When you sell access to a database, one price per seat is honest. The tenth heavy user and the tenth light user cost you roughly the same to serve, because your marginal cost per user is close to zero. So you pick a number, you say “unlimited,” and the buyer loves the predictability. Everyone wins.
AI products break the assumption underneath that model. Your revenue per user is still fixed by the subscription, but your cost per user is now genuinely variable, because it scales with how much they actually use the product. Every prompt is a real bill from OpenAI or Anthropic. The heavy user does not just engage more, they cost more, in direct proportion.
That is the structural flaw. Flat pricing pairs fixed revenue with a cost that used to be flat and is now variable. The two halves no longer match, and no amount of growth fixes a per-unit mismatch. You can scale a healthy flat-price business to the moon. You cannot scale your way out of selling something for less than it costs to deliver.
The numbers: 80-90% SaaS margin vs ~52% AI margin
The gap shows up in the margins. Traditional B2B SaaS runs 80-90% gross margins. AI-first products run roughly 50-60% on the same revenue, because inference is a direct variable cost sitting against every dollar you collect.
It gets more specific. Per ICONIQ’s 2026 State of AI survey, AI product builders expect average gross margins around 52% in 2026. And inference cost alone can eat roughly 23% of revenue for AI-first products. Nearly a quarter of every dollar is gone before you pay a single engineer, landlord, or ad.
Sit with the arithmetic for a second. In a flat-price AI business, your unit margin is 52% on average, and averages hide the users who are underwater. The whole point of a flat price is that it does not move. The whole problem is that your cost does. As the CFOs writing about this have started to warn, an AI feature bolted onto a classic subscription can quietly pull your gross margin down without anyone noticing on the revenue line.
Why a few users break the whole plan
Here is the part flat pricing never had to survive before: the distribution of usage in an AI product is brutally uneven, and your price is not.
On a flat plan, every user pays the same. But usage follows a long tail. Most users are light, a handful are heavy, and the heavy ones are not a little heavier, they are ten or twenty times heavier. When a normal SaaS user gets ten times more engaged, that is a success story. When an AI user gets ten times more engaged, that is ten times the token bill against the exact same subscription revenue. The person you would normally celebrate is the person setting your margin on fire.
Autonomous agents make this worse, and they make it new. A human power user is bounded by hours in the day. An agent your customer wires up against your API is not. It can loop, retry, and fan out thousands of calls while the account still pays one flat monthly fee. Flat pricing was designed for humans clicking buttons. It has no answer for a script that never sleeps sitting behind a $29 plan.
The math is simple and unforgiving. If your worst-case cost per user is unbounded and your revenue per user is fixed, your worst-case margin is unbounded loss. That is not a pricing tweak away from being fine. It is the model.
The three fixes: price from real cost, add usage or credits, and cap per user
You do not have to throw out subscriptions. You have to stop letting the subscription be the only lever. Three moves fix the structural problem, and the best AI pricing in 2026 uses some combination of all three.
1. Price the base from real cost, not vibes
Measure your actual AI cost per user for a real month, take the average, and set the base price at roughly 3-4x that number. That multiple is what holds a 60-70% gross margin after you account for the users above average. Pricing from a competitor’s number is guessing, because their model mix and prompt design are not yours.
2. Add a usage-based or credit component on top of the base
A pure usage model matches revenue to cost perfectly but scares buyers who want a predictable bill. So do not go pure. Keep a base subscription with a generous allowance, then charge usage or credits once a user goes past it. Normal users get the flat, predictable experience they wanted. Heavy users start paying in proportion to what they cost, which is exactly the alignment flat pricing was missing.
3. Put a spending cap on every plan
Even with a smart base and usage on top, put a hard per-user cap on each plan. When a user hits it, downgrade them to a cheaper model, gate the feature, or push an upgrade. A cap converts your worst case from “unbounded loss” into “a number you chose.” That is the single most important property a pricing model for an AI product can have.
How to actually price from cost
The mechanics are less mysterious than they sound. You need one month of honest data, then one decision per plan.
Step one, measure. For a full month, record the AI cost every user generates, tagged by their plan. A month matters because usage is spiky and a single week will lie to you.
Step two, aggregate by plan and read the shape, not just the average:
SELECT plan,
COUNT(*) AS users,
AVG(cost_usd) AS avg_cost,
PERCENTILE_CONT(0.90) WITHIN GROUP (ORDER BY cost_usd) AS p90_cost,
MAX(cost_usd) AS worst_case
FROM user_ai_cost_last_30d
GROUP BY plan;The average tells you where to set the base price. The p90 and the worst case tell you where to set the cap. If your p90 cost is already close to your price, the plan is fragile and one busy month tips it negative.
Step three, set the number. Take the average cost per user on that plan and multiply by 3-4x for the base price. Set the cap somewhere between the p90 and the point where the user turns unprofitable, so almost everyone is untouched and only the extreme tail gets gated. That is the entire method. The hard part was never the arithmetic, it was having trustworthy per-user cost in front of you in the first place.
Doing it without a data project
You can build the month of per-user cost tracking, the per-plan rollups, the pricing spreadsheet, and the caps yourself. It is a real data project, and it goes stale every time a provider reprices. Or you can wrap your LLM client with Weckr and get all of it as a byproduct of the calls you already make:
import { Weckr } from '@weckr/sdk'
const wk = new Weckr({
apiKey: process.env.WECKR_API_KEY,
plans: { free: 0, starter: 9, pro: 29, business: 99 },
})
// Per-plan spending caps are configured in the dashboard, not in code.
await wk.chat(openai, {
model: 'gpt-4o-mini',
messages,
userId: user.id,
feature: 'ai-summary',
plan: user.plan,
})Every call records cost and margin per user and per feature, calculated server side from current pricing so it never drifts when a provider changes rates. A pricing-recommendation view reads your real usage and suggests what to charge per plan, so you are setting the base from data instead of a hunch. And per-plan spending caps, which you set in the dashboard, do the third fix for you: when a user crosses the ceiling on their plan, Weckr can block the call or downgrade the model, turning your unbounded worst case into the number you set.
See it on seeded data, no signup, at useweckr.com/demo. The free tier covers 50k requests a month and Pro is $49/mo, so most teams can validate their pricing before they pay anything.
FAQ
Why doesn't flat subscription pricing work for AI products?
Because a flat subscription fixes your revenue per user while AI inference cost scales with how much each user consumes. A traditional SaaS product costs about the same to serve whether a user is heavy or light. An AI product does not, so a few heavy users can cost far more than they pay and quietly erase your margin.
What pricing model should an AI SaaS use instead of flat subscriptions?
Measure your real cost per user first, then do one or more of three things. Set the base price at roughly 3-4x average cost per user to hold 60-70% gross margin, add a usage-based or credit component on top of the base, and put per-user spending caps in place so heavy users get gated or downgraded instead of silently draining margin.
How do I stop a few power users from destroying my AI margins?
Put a per-user monthly spending cap on each plan. When a user hits the cap, downgrade them to a cheaper model, gate the feature, or prompt an upgrade. This turns your worst-case cost from unbounded into a known ceiling, which is the whole point of pricing an AI product.
Should I charge usage-based pricing for my AI features?
A pure usage-based model aligns revenue with cost perfectly but scares off buyers who want a predictable bill. The pragmatic answer for most AI SaaS is a hybrid: a base subscription that covers a generous allowance, plus usage-based or credit pricing once a user goes past it. You keep predictability for normal users and protect margin from the heavy ones.
How do I know what to charge per plan to stay profitable on AI?
Track your actual AI cost per user for at least one full month, take the average per plan, and set the price at roughly 3-4x that number for a 60-70% gross margin. Do not guess from your competitors, because your model mix and prompt design make your costs specific to you. Weckr computes this per-plan cost and recommends a price automatically.
Price from real per-user cost, not a guess
Flat, unlimited pricing feels safe because it is familiar. On an AI product it is the opposite of safe: it caps your revenue and uncaps your cost, and it hands your worst outcome to whichever user or agent decides to push hardest. The fix is not clever copywriting on the pricing page. It is pricing from real per-user cost, adding usage on top, and capping the tail.
All three start with seeing your actual cost per user, which is exactly what useweckr.com/demo shows you. Click through the seeded dashboard, watch the pricing-recommendation view suggest a number from real usage, and set caps that keep your heaviest users from quietly deciding your margin for you.