How to Price AI Features in SaaS When Every Prompt Costs You Money

‍

You've built an AI feature that works. Users love it. Now comes the hard part: how much do you charge?

If you run a subscription business, the old pricing playbooks don't fit anymore. AI changes everything about software economics. Every prompt costs you money. Every generation hits your margins. The more successful your product becomes, the more expensive it gets to run.

This guide breaks down what actually works for AI pricing right now, the mistakes that kill margins, and how to build pricing that scales with your business.

Why AI Pricing Is Different

Traditional SaaS has gross margins between 80-90%. Build it once, sell it infinitely. AI doesn't work that way.

Every time someone uses your AI feature, you pay for it. API calls. Compute. Storage. Safety layers. It adds up fast.

Think about it like a restaurant versus a consulting firm. The consulting firm has mostly fixed costs. One more client doesn't change much. But a restaurant? Every meal requires ingredients. Price the menu wrong and you lose money even when you're packed.

That's exactly what happens with AI. Two customers on the same $50/month plan can cost you wildly different amounts to serve. Your power user who sends 500 prompts a day? They might cost you $30/month. A casual user? Maybe $0.50.

So what margins should you actually target?

These numbers are lower than traditional SaaS. That's the reality you need to plan for.

And here's what makes this even trickier: costs are dropping fast. According to Ramp's data, the average cost per million tokens fell 75% in just one year. From $10 to $2.50. That sounds great until you realize your competitors can undercut you just as quickly.

So how do you price something when the underlying costs keep shifting? You build flexibility into your model from day one. More on that later.

The Seven Pricing Models That Actually Work

No single "right" answer exists here. The best model depends on your product, your customers, and your stage. But you need to understand all seven before you pick.

Usage-Based Pricing

Your customers pay for what they use. Per token, per API call, per generation.

This is how the big AI providers price their APIs:

Sources: OpenAI, Anthropic, Google

When does this make sense for your product? If you sell to developers who understand consumption metrics. If your cost-to-serve varies wildly between customers. If you want tight control over margins.

But here's the problem. Non-technical buyers don't understand tokens. They can't budget for something they can't predict. And one viral use case can crater your margins overnight.

Seat-Based Pricing With AI Tiers

Traditional per-user pricing, but with AI access allocated per seat.

GitHub Copilot does this well. $10/month for individuals. $19/user/month for Business. Notion bundles AI into their $20/user Business tier.

This works when AI enhances what humans do rather than replaces them. Collaboration tools. Productivity software. Anything where value scales with team size.

But think about this: what happens when your AI gets good enough that customers need fewer seats? You've accidentally created a situation where success limits your revenue.

Outcome-Based Pricing

Customers pay when AI delivers a specific result. Not per attempt. Not per message. Only when it works.

Intercom's Fin charges $0.99 per resolved conversation. If the AI doesn't resolve the issue, the customer doesn't pay.

This transforms how you sell. Instead of "how much does your AI cost?" the question becomes "how much do you currently pay per support resolution?" If a human agent costs $5-10 per ticket and Fin costs $0.99, the math is obvious.

The catch? You need clear success metrics. Attribution gets messy when multiple tools touch the same outcome. And your revenue becomes highly variable.

Hybrid Models

A monthly fee plus consumption charges. This is where most AI companies land.

Cursor does this at $20/month for Pro, which includes $20 worth of compute. Their Ultra tier at $200/month gives you 20x more. You get predictable revenue from subscriptions. You capture upside from heavy users through consumption. It works for almost any AI product.

Credit Systems

Customers buy credits upfront and redeem them for features. Midjourney does this with "Fast Hours." Runway charges credits per second of video.

Credits work well when you have diverse features with different cost profiles. But CFOs hate ambiguity. "What exactly is a credit worth?" can kill enterprise deals if you can't answer simply.

Freemium With AI Upsells

Basic AI free. Advanced features paid.

ChatGPT does this. Free tier with limits. Plus at $20/month. Pro at $200/month for power users.

This works if you have viral loops and enough funding to sustain free users. But be careful what habits you train. Give away too much and you'll struggle to monetize later.

Flat-Rate Unlimited

Fixed monthly fee. Usually with rate limits hidden in the fine print.

ChatGPT Plus at $20/month uses this for simplicity. But without careful capacity planning, your heaviest users will cost you more than they pay.

How to Calculate Your True Costs

You can't price intelligently without knowing your unit economics. Here's the formula:

Cost = ((input_tokens × input_rate) + (output_tokens × output_rate)) / 1,000,000

Quick rule of thumb: one token equals roughly 4 characters or three-quarters of a word.

Let's say you use GPT-4o-mini for email generation. Average email is maybe 600 tokens total. Your direct API cost? About $0.0006 per draft. Add 30% for infrastructure overhead and you're at roughly $0.0008.

That sounds trivially cheap. But multiply it by thousands of users who generate dozens of drafts daily. It adds up.

And that's just the obvious costs. What about the hidden ones?

Prompt creep. That 500-token system prompt you wrote? It gets sent with every single request. Long prompts plus verbose outputs can 3-5x your token usage.

Context stuffing. RAG implementations that pack the context window with retrieved documents are expensive by default.

Model overkill. "Always use GPT-4" is rarely optimal. Many tasks work fine with cheaper models.

Multi-step workflows. An AI feature that plans, researches, drafts, and edits might make 4-5 LLM calls per user action. Multiply accordingly.

Once you know your costs, work backward:

Minimum Price = Raw AI Cost / (1 - Target Margin)

If your raw cost is $0.80 per 1,000 calls and you want 75% margin:

• $0.80 / (1 - 0.75) = $3.20 per 1,000 calls
• Round to $3.50 and build in rate limits

What's Actually Working Right Now

Theory is useful. But what are successful companies actually doing?

The $20/Month Anchor

Something interesting happened. $20/month became the default for "AI productivity enhancement."

ChatGPT Plus. Claude Pro. Perplexity Pro. Cursor Pro. All exactly $20/month.

Why? It's close enough to Netflix to feel accessible. High enough to signal quality and filter out tire-kickers. If you're not sure where to price, $20/month is a safe anchor for individual users.

A second cluster exists at $10-12/month for entry-level offerings. GitHub Copilot Pro sits here.

At the premium end, $200/month has emerged for power users who need unlimited access. ChatGPT Pro lives at this price point.

How GitHub Copilot Structures Tiers

Their pricing follows a thoughtful approach:

Notice what they don't do? They don't charge per completion or per line of code. For a tool where AI helps you write code faster, usage-based pricing would create friction exactly when you want users to experience value.

How Notion Moved From Add-On to Bundled

Notion used to sell AI as an $8-10/user/month add-on. As of May 2025, according to reports on their pricing changes, full AI access requires the $20/user Business tier.

Why the change? As AI becomes expected in productivity software, a separate line item makes you look more expensive than competitors who bundle it. This is something to watch for in your own market. Are competitors bundling features you charge extra for?

How Intercom Made Outcome-Based Work

Intercom's Fin charges $0.99 per resolved conversation. Base platform costs ($29-132/seat) are separate.

This aligns incentives perfectly. Intercom gets paid more when Fin gets better at resolving issues. Customers don't pay for failures. The ROI conversation writes itself.

Could outcome-based pricing work for your product? Ask yourself: can you clearly define success? Can you measure it objectively? If yes, this model deserves serious consideration.

The Mistakes That Sink AI Pricing

Certain failure patterns show up repeatedly. Avoid these.

You Treat AI Like Just Another Feature

The most common mistake. "AI included, unlimited prompts" added to your existing $40/user plan because it seems simpler.

What happens next? A subset of users turn out to be power users. Thousands of prompts monthly. Each one costs you money. Suddenly you're losing money on your most engaged customers.

The fix? Either cap AI usage within standard seats. Or create an AI-specific add-on. Or build a hybrid model with metered usage.

Bill Shock

Pure pay-as-you-go sounds fair. Customers pay for what they use. But then they get their bill.

"Your AI usage this month: $847."

Nobody wants that surprise. Even if pricing was clearly explained, unpredictable bills create support nightmares and churn.

Users also get what's called "meter anxiety." They subconsciously pull back from features. They stop experimenting. They never find the value.

The fix? Hybrid pricing with bundled quotas plus clear overage rates. Usage dashboards. Budget caps and alerts. Make bills predictable.

Technical Pricing for Non-Technical Buyers

"$0.002 per 1,000 input tokens, $0.008 per 1,000 output tokens, plus $0.0001 per embedding query."

If you sell to developers, fine. If you sell to marketing managers who want help with emails, you've lost them.

The fix? Price on value your customers understand. Documents processed. Tasks completed. Conversations handled. Transform tokens into outcomes.

No Customer-Level Cost Tracking

Two customers on identical $500/month plans. Customer A uses AI heavily and costs you $350/month to serve. $150 margin. Customer B barely touches AI and costs $30/month. $470 margin.

Customer B is 3x more valuable. But without customer-level tracking, they look identical in your dashboard.

The fix? Instrument your product to track AI costs per customer from day one. You can't make smart decisions without this data.

Too Much in Free Tiers

Generous free tiers can drive adoption. But they also train expensive habits. Once users expect AI for free, that's painful to change.

The fix? Free tier allowances should reflect what you can sustainably handle while still showing value. Better to have users want more than to set expectations you can't maintain.

No Pricing Flexibility

API costs fell 75% last year. They could fall more. Or they could spike if you need to switch providers.

Multi-year fixed pricing for AI features is risky. So is building your position around a cost structure that might not exist in 18 months.

The fix? Pass-through clauses in enterprise contracts. Price guarantees limited to 12 months. Model abstraction so you can switch providers while you maintain service levels.

Pricing by Stage

What works at $5K MRR fails at $500K MRR.

Under $20K MRR

You're optimizing for learning, not perfect revenue.

• Start pricing conversations immediately. Wrong pricing at launch can mean negative margins before you can correct.
• Keep it simple. 2-3 tiers maximum. You don't have enough data for sophisticated usage models yet.
• Don't give pilots away free. That trains bad expectations. Use "early-bird" pricing instead.
• Target 60-70% gross margins from day one. Track cost per customer religiously.
• If you're bootstrapped, prioritize paid customers immediately. Tiered pricing attracts customers while it generates cash.

$20K-$200K MRR

You have customers and data. Now get more sophisticated.

• Move toward hybrid models. You have enough data to understand consumption patterns.
• Add tiers deliberately. Successful SaaS companies raise prices several times during growth. 10-15% each time. Major feature launches or flat conversion rates are good triggers.
• Build robust usage tracking. You can't do sophisticated pricing without detailed telemetry.
• Segment enterprise and self-serve. Enterprise expects multi-year agreements and 20-40% committed-use discounts. Self-serve wants simplicity and monthly billing.

Above $200K MRR

Pricing becomes a strategic lever.

• Consider outcome-based components. You have enough success data to tie pricing to results.
• Invest in pricing infrastructure. Real-time metering. Sophisticated billing. Usage dashboards. Budget controls. Manual tracking doesn't scale.
• Watch for competitive pressure. As AI costs fall, competitors will undercut you. Defend by moving up the value chain or building switching costs.

B2B vs. B2C

Different buyers need different approaches.

B2B

Business buyers make rational decisions. They need to justify purchases to procurement and finance.

Price based on value, not affordability. If your AI saves 10 hours of analyst time weekly at $75/hour, that's $3,000/month in value. A $500/month price captures only a fraction. You have room to be confident.

Expect negotiation on large deals. Budget 20-40% flexibility into list prices.

Provide ROI justification materials. Calculators, case studies, clear metrics. Business buyers need ammunition for internal approvals.

Push for annual contracts. Monthly billing creates constant renewal risk. 15-20% discounts for annual commitment are standard.

B2C

Consumer buyers make emotional decisions. They compare you to everything else competing for their attention.

Price sensitivity is high. Consumers expect free options and will search for them.

Decisions are impulsive but so are cancellations. Make signup frictionless. Expect churn if value isn't immediately obvious.

Psychological pricing works. $19.99 converts better than $20. $9/month converts better than $10/month.

Reference pricing matters. Consumers compare your AI tool to Netflix ($15) or Spotify ($10), not to the cost of hiring someone.

SMB

Small and medium businesses are a hybrid. More price-sensitive than enterprises. More rational than consumers.

They prefer pay-as-you-go because they can't commit to annual contracts without proven ROI. Self-serve onboarding is critical. They don't have time for lengthy sales processes. Make pricing clear on your website and your product usable without talking to anyone.

Global Pricing

If your product works globally, uniform pricing leaves money on the table.

Companies that do regional pricing often see higher growth. Purchasing power parity adjustments improve conversion significantly in emerging markets.

Typical discount ranges from US pricing:

Start simple with currency conversion and smart rounding (€19 instead of €18.73). As you scale, add full regional adjustments with local payment methods.

Watch for arbitrage. VPN users who buy at Indian prices while they use your product in the US is a real problem. Payment verification and usage monitoring help manage this.

Put It Together

AI pricing isn't something you solve once. It changes with your product, your market, and your customers.

The founders who get this right share certain habits.

They start with clear unit economics. Before any price is set, they understand costs at a granular level. Not just API calls. The full loaded cost to serve each customer.

They instrument everything from day one. No intelligent pricing decisions happen without data on usage patterns, cost drivers, and willingness to pay.

They align their success with customer outcomes. The strongest pricing models are ones where you make more money when your customers succeed.

They stay flexible. Pricing decisions made today will need to change. They build that expectation into contracts and thinking.

They resist the temptation to give everything away. Free tiers can drive adoption. But customers trained to expect AI for free create long-term monetization problems.

The core insight? As AI costs fall, value-based and outcome-based pricing become more important. Companies that charge per token face relentless pressure. Those that charge for business results can maintain pricing power regardless of what happens to underlying model economics.

Start simple. Measure everything. Iterate quickly. What pricing model makes the most sense for where your business is today?

‍