Pricing AI Products

Here is the conclusion first, because it will save you from the most common mistake founders make: AI pricing is structurally different from SaaS pricing because your costs scale with usage, not just users — and almost every founder prices too low.

Traditional SaaS has near-zero marginal cost per user. Once the software is built, adding the 10,000th customer costs almost nothing. This lets SaaS companies build at a price that grows linearly with users while costs stay roughly flat. Gross margins of 80%+ are the norm.

AI products don’t work this way. Every query, every generation, every inference call has a real dollar cost attached to it. A user who runs your AI coding assistant for 8 hours a day costs you 50x more than a user who logs in twice a week. But they both pay you $20/month. If you haven’t priced for this, you will find out at scale — when your ARR is growing but your gross margin is contracting toward zero.

The founders who get pricing right early build sustainable businesses. The ones who get it wrong raise more money to fund their own cost spiral and eventually either raise prices aggressively (losing customers in the process) or quietly die.

Let’s build a framework for getting it right.

1. The AI Cost Structure Problem

In SaaS, your cost structure looks like this: fixed infrastructure that scales slowly + support + headcount. Adding more users barely moves the cost needle.

In AI products, your cost structure has an additional line item that traditional SaaS doesn’t: inference cost, and it scales directly with usage intensity.

The math is straightforward. GPT-4o charges roughly $2.50 per million input tokens and $10 per million output tokens. A product that generates 2,000-word responses might be spending $0.02–$0.05 per generation. That sounds trivially small. But if your power users generate 100 responses a day, you’re spending $2–5 per user per day, or $60–150 per user per month. Your $29/month plan is deeply underwater on those users.

This creates what I call the heavy user problem: your best, most engaged users — the ones who love your product and would tell their friends — are the ones who cost you the most. In traditional SaaS, your power users are your best customers. In AI, they can be your biggest liability if you haven’t priced for them.

The specific numbers vary by model and use case, but the structural problem is the same everywhere:

• Voice AI apps burn significant compute per minute of audio
• Image generation apps have hard per-image costs
• Code assistants with long context windows carry large per-session token costs
• Agentic products that chain multiple LLM calls can spend $0.10–$1.00 per completed task

Before you set any price, you need to know your cost per active user at 50th, 90th, and 99th percentile usage. Most founders know their average. The average will lie to you.

2. Pricing Models and When Each Works

Per-Seat SaaS

The classic model: charge per user per month, regardless of how much they use it.

When it works: When usage is relatively uniform across users, or when your AI feature is embedded in a broader workflow tool where users are paying for the whole package, not just the AI component. GitHub Copilot ($10/month Pro, $19/user/month Business) largely follows this model — as of June 2026 it added AI Credits for chat and agent features, but code completions remain on a per-seat subscription. The variance in how much developers type isn’t extreme enough to blow up the model, and the distribution of usage is fairly predictable.

When it breaks: When you have high usage variance — a segment of power users who hammer the product 10x more than average. Per-seat pricing in this scenario means your best customers are subsidized by your casual ones, and any shift toward more engaged usage tanks your margins.

The fix if you want to stay per-seat: Build usage floors into your model assumptions and price above them. If your P90 user costs you $15/month in compute, don’t charge $19/month with a thin margin. Charge $49/month and accept that casual users are highly profitable subsidizers.

Usage-Based Pricing

Charge per query, per generation, per API call, per action completed.

When it works: Developer-facing products where users understand and expect consumption pricing. Products where value and cost scale together naturally — e.g., a document extraction tool where 1,000 pages processed = 1,000 pages of value delivered. Products with high variance in usage where per-seat would be deeply unfair (some users need 10x the resources of others).

When it breaks: For consumer and SMB products where unpredictable bills are anxiety-inducing. Perplexity learned this: their early usage-based model caused hesitation among casual users who didn’t want to think about costs per query. They moved to a flat subscription model to reduce friction. Usage-based pricing also creates revenue unpredictability on your end — a down month in usage from a single large customer can crater MRR.

The fix: Usage-based works best when you add a floor (monthly minimum) and ideally a ceiling (cap or overage model) so both sides know what to expect.

Outcome-Based Pricing

Charge per result: per lead generated, per bug fixed, per deal closed, per hour saved.

When it works: When the outcome is clearly measurable, the value is high relative to cost, and the customer is sophisticated enough to trust the measurement. Sales AI tools that charge per qualified meeting booked. Legal AI that charges per contract reviewed. Recruiting tools that charge per successful placement.

When it breaks: When outcomes are hard to attribute to your product specifically (did the AI close the deal, or the salesperson?), when customers dispute measurement, or when outcomes vary wildly (one customer’s “hour saved” is worth $50, another’s is worth $500 — do you charge the same?).

Outcome-based pricing is the highest-margin, highest-alignment model when it works. It’s also the hardest to sell and implement. Most early-stage founders should get it working on per-seat or usage-based first, then layer in outcome-based elements later as proof points accumulate.

Hybrid: Seat + Usage Cap

The most common real-world model for AI SaaS: charge per seat, include a monthly usage allowance, charge overages or hard-cap beyond it.

This is how most successful AI tools actually work in practice. Notion AI includes a credit bundle in paid plans and charges overages. Many AI writing tools give 10,000 words/month and charge beyond that. Cursor bundles fast requests per month with their subscription.

Why it works: Predictable revenue from the seat charge. Protection from the heavy user problem via the cap. Upside from power users who pay for overages. Psychological framing of “included usage” that feels generous rather than restrictive.

The tradeoff: Customers hate overage charges if they’re opaque. Be explicit about what’s included, what triggers overages, and what overages cost. The worst thing you can do is have a “soft” usage limit that some customers hit and some don’t, with no clear policy — it creates customer service issues and perceived unfairness.

3. Calculating Your Floor Price

Before you talk about value-based pricing or market positioning, you need a hard floor — the price below which you’re guaranteed to lose money on meaningful usage.

The framework:

Monthly Cost Per User = 
  (Average inference cost per session × avg sessions/month)
  + (Infrastructure cost / active users)
  + (Support cost / active users)

Floor Price = Monthly Cost Per User × (1 / target gross margin)

Example for a code review tool:

• Average inference cost per review: $0.08
• Average reviews per user per month: 40
• Monthly inference cost: $3.20/user
• Infrastructure and support allocation: $1.50/user
• Total cost per user: $4.70/month
• Target gross margin: 70%
• Floor price: $4.70 / 0.30 = $15.67/month

That’s your floor — the minimum price to hit 70% gross margin on an average user. But remember, this is the average user. Your P90 user might do 150 reviews a month. Your floor for that user is $58.50/month. If they’re all on your $19/month plan, you’re losing money.

What to do with this number:

Price above it — significantly above it. The floor is not the target price. It’s the guardrail. Your target price should be anchored to value delivered, not cost structure. But if value-based pricing puts you at $19/month and your floor is $22/month, you either need to reduce costs, raise prices, or kill the product.

Run this calculation for P50, P90, and P99 usage before you publish any pricing. Then decide what usage tier your plan implicitly assumes, and make sure you have mechanisms (caps, overage pricing, terms of service) to prevent the P99 user from destroying your margins.

4. Value-Based Pricing: Anchor to Outcomes, Not Costs

Costs tell you where to set the floor. Value tells you where to set the ceiling — and almost always, the ceiling is much higher than founders think.

The question to ask is: what is this product worth to the customer in concrete terms?

A recruiting tool that saves a recruiter 5 hours per week is worth $200–500/month per seat — at minimum. If you charge $49/month, you’re leaving 80% of the value on the table.

A code review tool that catches bugs before production — bugs that would cost 2–10 hours of engineering time each — is worth hundreds of dollars per developer per month. Charging $19/month because “Copilot is $10” is competitive anchoring to the wrong product.

The anchor interview: Talk to 10 customers and ask them: “If this product disappeared tomorrow, what would you do instead? How long would that take? What would it cost you?” The answers will calibrate your value estimate. When a customer says “I’d have to hire someone,” you’re in high-value territory. When they say “I’d just use ChatGPT directly,” you have a differentiation problem.

Pricing anchors to use:

• Replacement cost: What does it cost to do this without your product? (Hire, outsource, manual process)
• Time value: How many hours does your product save? × the user’s hourly rate
• Outcome value: What does a successful outcome generate? (Revenue, risk avoided, deals won)

Price at 10–20% of the value you deliver. If you’re capturing less than that, you’re either underpriced or the customer doesn’t believe your value claim — which is a sales problem, not a pricing problem.

5. Freemium vs. Free Trial vs. Paid-Only

This is a common trap: defaulting to freemium because “it worked for Slack” without thinking through the AI-specific economics.

Freemium for AI products is dangerous without hard usage limits. If your free tier offers substantive AI usage with no cap, you will attract large numbers of users who never convert and cost you real money in inference. The conversion math has to work: if your free-to-paid conversion rate is 3%, each free user costs you N months of inference before they convert or churn. Model this before you ship freemium.

What works:

• Freemium with aggressive limits: Give free users enough to experience the product (5 generations, 3 documents, 10 queries) but not enough to replace paid. Cursor does this well — you get a trial of fast requests, then you hit a wall.
• Free trial (time-limited, full features): 14-day trials with access to the real product. This is usually better for AI tools than freemium because it filters for actual buyers and doesn’t subsidize permanent free riders.
• Paid-only: Valid for developer tools and enterprise products where the audience understands and expects paid-only access. Removes the conversion funnel but also removes the cost leakage.

The case against freemium: Support costs. Free users submit support tickets at roughly the same rate as paid users, but generate no revenue. For AI products where free-tier users are often students, researchers, or people who never intend to pay, the support-to-revenue ratio is brutal.

6. Pricing by Segment

The same product at the same price will convert very differently across segments.

Self-serve / individual ($0–$50/month): Price point is mostly psychological. Under $20/month is an impulse buy. $20–50/month requires a clear value story. Above $50/month for an individual requires either significant demonstrated ROI or a power-user audience who bills by the hour and sees it as a professional tool. Keep friction low, onboarding automated, and support self-serve.

SMB ($100–$500/month): Now you’re selling to a budget holder who needs a business case. Pricing per seat × team size often works better than flat pricing. The value story needs to connect to business outcomes, not personal productivity. SMBs want to know: will this make my team faster? Will this reduce headcount? Prove one of those things.

Enterprise ($1,000+/month): Custom pricing, contracts, procurement cycles, and security reviews. The list price becomes a starting point for negotiation. The actual value drivers are reliability, compliance, SLAs, and integrations — not AI quality. At enterprise scale, you also need to worry about volume pricing: a 500-seat deployment should cost less per seat than a 10-seat deployment, but not so much less that you’re giving away the product.

One important principle: don’t let enterprise deals set your self-serve price anchor. Early enterprise customers will sometimes push for custom deals that are well below what self-serve customers will pay. If you get locked into $5/seat/month enterprise contracts, you’ll have a hard time charging $15/month to self-serve customers who discover the gap.

7. Common Pricing Mistakes

Exposing per-token pricing to users. Users don’t think in tokens. Showing “you have 50,000 tokens remaining” creates anxiety and confusion. Abstract it into something meaningful: generations, documents, hours of usage, requests. The mental model has to match how users experience the product.

Pricing too low to cover support. Support at scale is expensive, and AI products tend to attract support-heavy users because the product behavior is less predictable than traditional software. A plan priced at $9/month that generates one support ticket per month is unprofitable before you add any other costs. If you’re competing on price, you must also compete on reducing support burden — which usually means much better onboarding, not cheaper plans.

No usage limits on the free tier. Covered above, but worth repeating: free without limits is not a marketing strategy, it’s a slow burn on your compute budget. Every week you run freemium without limits is a week of burning investor capital to subsidize non-customers.

Anchoring to competitors without validating the value delta. “Copilot is $10/month so we should be $15/month” ignores whether your product is better, worse, different, or in a different category. Competitor pricing tells you what the market has accepted before — not what your specific value proposition is worth.

Flat pricing for a product with highly variable costs. If a customer can generate 100x more value (and cost you 100x more) at the same price point, you have a pricing design problem that will eventually become a business problem.

8. How Real Products Price: Cursor, Perplexity, Notion AI, GitHub Copilot

GitHub Copilot ($10/month Pro, $19/user/month Business; new Pro+ at $39 and Max at $100): Per-seat subscription with AI Credits for chat and agent features (added June 2026); code completions remain unlimited. Works because inference per autocomplete is cheap (short context, fast model), usage variance is bounded by developer working hours, and the value anchor is clear — developer productivity is worth far more than the base subscription price.

Cursor (~$20/month for Pro): Hybrid. Pro includes unlimited Auto mode requests and a credit pool for manually selecting premium models. Overages are available at API rates. Smart design: the free cap creates conversion pressure, the overage model monetizes power users, and the flat fee creates predictability.

Perplexity ($20/month for Pro): Flat subscription with usage limits on the most expensive features (Pro search). Evolved away from pure usage-based to reduce friction. Works because the free tier is genuinely useful (driving growth) while the paid tier unlocks meaningfully better performance. Clear free-to-paid upgrade path.

Notion AI (bundled into the Business plan at $20/member/month; the standalone add-on was retired in 2025): Intelligent because it’s layered on top of an existing product relationship. The AI pricing piggybacks on team size — large teams pay more — which correlates roughly with value delivered. The risk is that teams see it as an AI feature they can do without rather than core infrastructure.

The pattern across all of them: flat or near-flat consumer pricing with usage guardrails on the expensive features. None of them expose raw compute costs to users. All of them have some form of usage tier that protects against the P99 heavy user.

9. How to Raise Prices Without Losing Customers

You will almost certainly need to raise prices. The AI cost landscape is still volatile, your initial pricing was probably based on limited data, and your product is worth more than you thought when you launched. Here’s how to do it without a revolt.

Grandfather existing customers aggressively. Lock in current customers at their current price for 12 months or longer. This converts “price increase” into “here’s why new customers pay more than you.” It’s a retention gift and a loyalty signal simultaneously.

Raise new customer prices first, existing last. New customers have no anchor to your old price. Test higher prices on new signups before rolling out to existing customers. If your new-customer conversion doesn’t crater, you have validation for a broader raise.

Add a tier, don’t just raise the price. If you’re raising from $29 to $49, add a $49 tier with clearly more value (higher limits, new features, priority support) before removing the $29 tier. Some customers will upgrade voluntarily. Others get a grace period before the existing tier disappears.

Be transparent about why. “Our infrastructure costs increased and we need to price for sustainability” is a legitimate, understandable reason. Founders who hide behind vague “new features!” justifications when everyone knows prices are going up damage trust. Just say the thing.

Give notice. 60–90 days for existing customers. 30 days minimum. Surprise price increases on subscription software feel like betrayal. Long notice periods feel like respect.

The Pricing Calculator Framework

Use this before you launch — or to audit your current pricing:

If your actual price is below your P90 floor price, you have a problem. If your actual price is below 10% of the value you deliver, you’re leaving money on the table. If there’s no usage mechanism to prevent P99 users from destroying your margins, you have a ticking clock.

Fill this in honestly. Most founders who do will either discover they’re underpriced, that they don’t know their costs well enough to price confidently, or both.

Pricing is not a launch decision — it’s an ongoing product decision. The founders who treat it that way, revisit it quarterly, and instrument their usage data from day one are the ones who build AI businesses that stay solvent at scale. The ones who set a price in week one and never revisit it are the ones who raise a Series A to fund a slow margin collapse.

Get the floor right. Anchor to value. Build in protection against heavy users. Then raise prices before you think you need to.