Vibe a Business (3/4): Pricing AI-Powered Products When Every User Costs You Money

This is Part 3 of How to Vibe a Business, a four-part series on the economics of building and selling AI-powered products as a solo developer. Part 1: The economics of one | Part 2: The token tax | Part 3: Pricing (you are here) | Part 4: Metrics after launch

For fifteen years, the default SaaS pricing model was “$X per user per month, unlimited usage.” It built Slack, Notion, Linear, and a generation of B2B companies. It works beautifully when the marginal cost of one more user doing one more action is close to zero — when your biggest expense is the engineer who shipped the feature, not the compute that runs it.

That model breaks the moment your backend calls a $15-per-million-token API.

In traditional SaaS, your power users — the ones who log in daily, who build workflows, who become internal champions — are your best customers. They drive expansion revenue, reduce churn, and generate word of mouth. In AI SaaS, those same power users can be your worst margin. A light user who pokes around twice a month might cost you $1.50 in API calls. A power user who runs your product hard every day can cost you $30 or more. Same $29/month subscription. Radically different economics.

Most solo builders discover this after launch, when the Anthropic or OpenAI invoice arrives and the number is larger than the revenue. If Part 2 showed you how to calculate your per-user AI costs, this post shows you how to price around them so you actually make money.

Why flat-rate pricing fails for AI products

Walk through the math with real numbers. Say you launch at $29/month — a common indie SaaS price point. Your median user costs $4 in API calls per month. That looks fine: $29 revenue minus $4 COGS leaves $25 of gross profit, an 86% margin. Ship it.

But the median hides the distribution. Your lightest 20% of users cost $0.50/month each. Your heaviest 20% cost $18/month. Your top 5% — the users who love your product enough to use it for two hours a day — cost $30 or more. You are paying those users to use your product.

In traditional SaaS, the cost curve is flat. One hundred users and ten thousand users run on the same servers, and the incremental cost per user is pennies. The pricing conversation is about value, not cost. In AI SaaS, the cost curve is linear (or worse). Every API call has a price tag. Every generation, every summary, every analysis hits an endpoint that charges by the token. Your COGS scales directly with engagement.

This inverts the fundamental SaaS dynamic. The behavior you most want — deep, daily product usage — is the behavior that destroys your margin. If you are building an AI product in 2026, do not launch with unlimited usage at a flat monthly price. You will subsidize your best customers with revenue from your worst ones, and when the product takes off, the subsidy will scale faster than the revenue.

Four pricing models that actually work

The industry is moving fast on this. As a16z noted in their analysis of AI business models, the economics of AI products are structurally different from traditional software — gross margins run 50-60% instead of 80-90%, and pricing has to account for that gap. Here are the four models that work, with honest tradeoffs for each.

1. Usage-based pricing

Charge per action or per output. The user pays for what they consume, and your revenue tracks your costs automatically.

Vercel charges per function invocation. OpenAI charges per token. Twilio charges per message. The model works cleanly when the “unit of value” is obvious — when users can easily understand what they are paying for and predict their bill.

Best for: Products where the value unit is clear (generations, analyses, translations, summaries) and users are comfortable with variable billing.

The downside: Unpredictable bills scare users away. “I don’t know what this will cost me” is a real conversion killer, especially for individuals and small teams. You also need billing infrastructure that meters usage in real time, which is non-trivial for a solo builder.

2. Credit-based pricing

Sell credits in fixed bundles that users spend on AI features. 100 credits for $19, 500 credits for $49. Each AI action costs a set number of credits depending on its complexity.

This is becoming the default pattern for AI-native products, and for good reason: it gives users cost predictability (they bought a fixed pack) while giving you revenue predictability (you sold the credits upfront). Midjourney popularized this model. Most AI image and video tools use it now.

Best for: Products with multiple AI features of varying cost, where you want to let users allocate their budget across actions.

The downside: Credit systems add UX friction. Users have to learn the credit economy, monitor their balance, and decide when to top up. If you set credit prices wrong, you either leave money on the table or frustrate users who feel nickeled-and-dimed.

3. Tiered pricing with usage caps

$19/month for 100 AI actions. $49/month for 500. $99/month for 2,000. This is the most common model for solo builders because it looks and feels like traditional SaaS to users — they pick a plan and pay a predictable amount.

The trick is setting the tiers so that each one is profitable at its cap. If your $49 tier includes 500 actions and each action costs you $0.08 in API, your AI cost at max usage is $40 — leaving you $9 of gross profit on a customer who is burning every available action. That is not a business. You need the cap low enough or the price high enough that even a power user at the limit gives you healthy margin.

Best for: Solo builders who want familiar SaaS packaging. Users get predictability, you get cost containment.

The downside: The tier boundaries create friction. Users approaching their cap either stop using the product (bad for retention) or feel pressured to upgrade (bad for satisfaction if the jump is too steep). Design the tiers so the natural upgrade feels like “I’m getting more value” rather than “I’m being punished for using the product.”

4. Outcome-based pricing

Charge for results, not inputs. Per resolved support ticket, per generated report, per qualified lead. As a16z argues, AI shifts value from “users” to “output” — and outcome-based pricing aligns your price with that shift.

Best for: Products where the output has a clear, measurable dollar value to the customer. If your AI tool saves a support team 10 hours per week, charging per resolved ticket is easy to justify.

The downside: Hardest to implement. You need to define “outcome” precisely, measure it reliably, and convince users that your measurement is fair. It also requires that outcomes are predictable enough for users to budget. But when it works, it has the highest perceived value — users pay for what they get, not what they use.

Model comparison

Model	Best for	Downside	Example
Usage-based	Clear value units, technical users	Unpredictable bills, metering complexity	OpenAI, Vercel
Credit-based	Multi-feature products, mixed usage	UX friction, credit economy design	Midjourney, AI image tools
Tiered with caps	Solo builders, SaaS-familiar users	Tier boundary friction, cap anxiety	Most indie AI SaaS
Outcome-based	High-value measurable outputs	Hard to define and measure, trust gap	AI support tools, report generators

If you are a solo builder launching your first AI product, start with tiered pricing and usage caps. It is the fastest to implement, easiest for users to understand, and gives you the cost containment you need while you learn your actual usage distribution. You can always migrate to credits or outcomes later once you have data.

The freemium trap for AI products

In traditional SaaS, free users cost approximately nothing to serve. A free Notion user sits in a database row, loads some static assets, and generates zero marginal cost. That is why freemium built Dropbox, Slack, and Figma — the free tier was essentially an unlimited marketing channel with near-zero COGS.

In AI SaaS, free users cost real money every time they use an AI feature. Run the numbers: 1,000 free users making 50 AI calls per month at $0.003 per call is $150/month just for the free tier. At 5% conversion to a $29/month paid plan, those 1,000 free users generate 50 paying customers and $1,450/month in revenue — of which the free tier just consumed 10%. Scale to 10,000 free users and you are spending $1,500/month to support users who may never convert.

If you offer a free tier — and you probably should for discovery and conversion — cap AI usage aggressively. Five to ten AI actions per day is enough for users to experience the value. More than that and you are subsidizing usage without a conversion path. Make sure your conversion math works: if your free-to-paid rate is below 3%, either improve the conversion funnel or tighten the free cap.

Use the CAC vs LTV calculator to model the free tier as an acquisition cost. Every dollar you spend serving free users is effectively CAC — and it needs to produce enough paid conversions at a high enough LTV to justify the spend.

How to set your actual price

Stop thinking about pricing as a creative exercise. For AI products, pricing is a math problem first and a value problem second. Here is the step-by-step:

Step 1: Calculate your per-user AI cost at median usage. Use the cost-per-unit calculator and plug in your actual API costs from Part 2. If you do not have real usage data yet, estimate conservatively — assume your median user runs 3-5 AI actions per day and size the token cost per action from your testing.

Step 2: Add infrastructure overhead. Hosting, authentication, payment processing, error monitoring, email. For a solo-built product, this typically runs $2-5 per user per month depending on your stack. Do not forget Stripe’s 2.9% + $0.30 per transaction — on a $19/month product, that is $0.85/month per user, which matters.

Step 3: Apply your target gross margin. If your gross margin is below 60%, you do not have a SaaS business — you have a services business with software wrapping. A16z’s analysis of AI companies shows gross margins of 50-60% for AI-native products, versus 80%+ for traditional SaaS. Target 65-70% to give yourself room. The markup vs margin calculator will convert between your cost markup and your gross margin so you do not confuse the two.

Step 4: That is your floor. Now price for value above it. The floor tells you the minimum price at which you do not lose money. The actual price should be set by the value you deliver, not the cost you incur. Use the value-based pricing calculator to model what your product is worth to customers — the time it saves, the revenue it generates, the cost it eliminates — and price at a fraction of that value. If your floor is $17 and your value-based price is $49, charge $49. If your floor is $17 and your value-based price is $15, you have a cost problem, not a pricing problem.

Worked example: an AI meeting summarizer

You have built an AI tool that joins video calls, transcribes them, and produces structured summaries with action items. Here is how to price it.

Per-user AI costs (median usage). Your median user has 12 meetings per month, averaging 30 minutes each. Transcription plus summarization costs roughly $0.27 per meeting. That is $3.20/month per user in API costs.

Infrastructure overhead. Hosting on Vercel, auth via Clerk, payments via Stripe, plus transactional email: $2.00/month per user at your current scale.

Total COGS per user. $3.20 + $2.00 = $5.20/month.

Target gross margin. 70%. Minimum price = $5.20 / (1 - 0.70) = $17.33/month.

Round up to $19/month for the base tier. Include up to 20 meetings per month (your median user uses 12, so this gives headroom without being unlimited). Run this through the pricing calculator to verify the margin holds.

Add a $49/month pro tier with up to 100 meetings per month. At 100 meetings, your API cost is $27.00 + $2.00 infra = $29.00 COGS. Gross margin: 41%. That is tight — but only your heaviest users hit the cap, and most pro users will land around 40-60 meetings, where the margin is healthier. If it gets too tight, raise the pro tier to $59 or lower the cap to 75 meetings. The math will tell you.

Add a free tier with 3 meetings per month. COGS: $0.81 + $0.50 (minimal infra) = $1.31/user/month. If your free-to-paid conversion rate is 5%, each paying user “costs” 19 free users in the funnel, or $24.89 in free-tier COGS. That is effectively your CAC for organic conversions — compare it against the LTV of a paying user ($19/month at 70% margin for 18 months = $239.40 LTV). The ratio works. Model it in the CAC vs LTV calculator to see the payback period.

The pricing decision you should make this week

If you are building an AI product and you have not set prices yet, do this:

Calculate your per-action API cost from real testing (not estimates, not the provider’s pricing page — actual costs from your logs).
Multiply by your expected median monthly usage to get per-user COGS.
Add infrastructure costs.
Divide by 0.30 to get the minimum price at 70% margin.
Set three tiers: free (aggressive cap), base (comfortable median), pro (generous but bounded).
Launch.

You can refine later when you have data. What you cannot do is launch at a flat $29/month unlimited and figure it out after your API bill exceeds your revenue. By then, the users who love your product the most are the ones costing you the most, and raising prices or adding caps retroactively is a much harder conversation than setting them correctly from day one.

The pricing calculator will run the full model for you: cost inputs, margin targets, and tier structure. Open it and put your numbers in before you put your price on your landing page.

Next in this series: Part 4 — The SaaS metrics that matter after launch, covering churn, MRR, and the dashboard you actually need as a solo builder.