Microsoft priced GitHub Copilot at $10 a month. The cost to serve ran roughly $30. The gap went unmeasured for two years, across 4.7 million subscribers.
Intercom priced Fin at $0.99 per resolved ticket and kept its margin. Same market, same models, opposite outcome.
The difference is not pricing talent. It is whether cost-to-serve data existed before the price was set. That is the whole framework in one sentence: price AI products from your cost-to-serve data up, not from competitor pricing down.
Intercom's $0.99 was not pulled from competitor analysis. It came from inference cost data, $0.15 to $0.85 per resolution depending on complexity, with enough volume for the law of large numbers to stabilize margin. Cognition's Devin charges $20/month plus per-compute-unit billing because autonomous coding sessions vary in cost by 100x. No flat price holds across that range.
The Margin Buffer Is Gone
Traditional SaaS has near-zero marginal cost per user. Add a customer and the infrastructure bill barely moves. That made it possible to price on value, competitive positioning, or gut feel and still hold 70-90% gross margins. The buffer forgave almost any pricing mistake.
AI products removed the buffer. Every interaction triggers inference, every inference costs money, and two customers on the same plan can differ in cost by 10-100x. A chat feature and a document analysis feature have entirely different cost profiles. A "free" feature nobody asked for might be quietly burning 40% of your compute.
The numbers on this are recent. ICONIQ Growth reports that scaling AI B2B companies averaged 52% gross margins in January 2026, up from 41% in 2024. Bessemer Venture Partners puts AI gross margins at 50-60%, against 70-90% for mature SaaS. For every $1M in AI product revenue, roughly $230,000 goes to inference alone, before engineering, sales, or support see a dollar.
| Dimension | Traditional SaaS | AI-First SaaS |
|---|---|---|
| Gross margin | 70-90% | 20-60% (avg 52%) |
| Marginal cost per user | Near zero | Variable, per-request |
| Cost predictability | High (fixed infra) | Low (model-dependent) |
| Cost variance across customers | Minimal | 10-100x range |
| Pricing model convergence | Settled (per-seat) | Unsettled (7+ models) |
Pricing an AI product requires cost-to-serve data at the customer level. Without it, margin calculations are fiction.
Cost Accounting Starts With the Trace, Not the Token
Most AI companies track the monthly OpenAI or Anthropic bill. Fewer track cost per customer. Almost none track cost per feature, per workflow, or per outcome.
The unit of AI cost accounting is the trace: a complete workflow execution from user request to final response. A single customer interaction might embed a query, retrieve context from a vector database, call a planner model, execute tool calls, generate a response, and run safety checks. That is six cost centers, often across multiple providers, in one interaction. Token math captures maybe two of them.
Current API prices show the range. GPT-4o costs $2.50/$10 per million input/output tokens, Claude Sonnet 4.6 costs $3/$15, and Claude Opus 4.6 costs $5/$25. A workflow that chains models, uses retrieval, and includes tool calls accumulates costs across all of them. Prompt caching cuts cached input costs by 90% and batch processing takes 50% off, but only if the architecture was designed for them.
Before setting any price, four questions need answers.
What does it cost to serve each customer per month? Not on average. Per customer, with variance, because averages hide the customers who cost more than they pay.
Which features drive the most cost? Knowing which feature costs what determines which features belong in which tier.
What does a single workflow execution cost? Not the LLM call alone, but the full trace: retrieval, tool calls, retries, safety checks.
And how does cost correlate with value delivered? Some high-cost interactions produce high value and some don't. That correlation determines whether outcome-based pricing is viable for your product.
If you can answer these four, you can price with confidence. If you can't, any pricing framework is built on assumptions that may be off by an order of magnitude.
Two Customers, One Plan, Different Businesses
A customer pays $100/month, costs $10/month to serve, and cost $300 to acquire. Payback is 3.3 months, three-year gross LTV is roughly $3,240, and LTV/CAC lands at 10.8x. Healthy.
Now take the same plan with heavier hands on it. A user who costs $80/month to serve stretches payback to 15 months and drops LTV/CAC to 2.4x. Same product, same price, different business. Per-customer cost data is what tells these two apart.
The model underneath is simple. Gross margin per customer equals revenue minus variable cost to serve, where variable cost includes LLM API costs across all models, embedding and retrieval infrastructure, tool call and external API costs, compute for self-hosted models, and a proportional share of vector databases and caching layers.
Bessemer sets the benchmark tiers. Fast-ramping AI "Supernovas" average about 25% gross margin early on. Steadier "Shooting Stars" trend closer to 60%. LLM-native companies holding around 65% while growing 400% year-over-year are the current ceiling.
A healthy AI product targets 60-80% gross margins. Below 50%, pricing or cost structure needs attention. The structural floor for AI companies will likely settle at 60-65%, below the 80%+ traditional SaaS achieves.
| Metric | Target | Why |
|---|---|---|
| Gross margin | 60-80% | Covers fixed costs and profit |
| LTV/CAC | 3x minimum | Validates acquisition economics |
| Payback period | Under 12 months | Runway sustainability |
| Cost variance | Measured, not averaged | Tiers should reflect reality |
| Inference as % of revenue | Under 23% | ICONIQ benchmark for scaling AI |
Seven Pricing Models, No Consensus
The market has not converged. A 2025 industry report found 92% of AI software companies now use mixed pricing models. Here is the map:
| Model | Example | Billing Unit | Works When |
|---|---|---|---|
| Input-based | OpenAI, Anthropic | Per token/call | Horizontal platform, technical buyers |
| Per-resolution | Intercom Fin ($0.99) | Resolved ticket | Binary outcomes, high volume |
| Per-lead | 11x ($5,000/mo) | Qualified lead | Measurable pipeline value |
| Hourly | Copilot for Security ($4/hr) | Compute hour | Bursty, unpredictable workloads |
| Outcome % | Chargeflow (25%) | % of value recovered | High-value, measurable outcomes |
| Hybrid seat+usage | GitHub Copilot ($19-$39/seat) | Seat + overages | Most B2B SaaS products |
| Bundled | Notion ($20/seat) | Feature-gated tier | AI as feature, not core product |
The right choice depends on your cost structure and how specific your use case is.
Input-Based Pricing
Tokens, API calls, compute time, storage. OpenAI and Anthropic use this model because they serve an entire market and cannot predict what gets built on their APIs.
It works when you're a horizontal platform, your customers are technical, and cost scales predictably with usage. The risk is behavioral. Clay discovered that when it introduced per-action pricing, users during onboarding chose to enrich 10 emails instead of 1,000. Not because 10 was enough, but because they were nervous about spending credits they couldn't predict. The onboarding trained users to be conservative instead of discovering value.
Outcome-Based Pricing
Resolved tickets, completed reports, qualified leads. Intercom Fin charges $0.99 per resolved conversation, 11x bills roughly $5 per qualified lead, and Chargeflow takes 25% of recovered chargeback value.
This works when the outcome is measurable and the problem is well-defined. The risk is cost variance per outcome. Some support tickets take 500 tokens to resolve and some take 100,000. If Intercom's average inference cost per resolution is $0.30, they keep $0.69 of gross margin. If a harder ticket requires multiple model calls and costs $0.85, margin drops to $0.14.
Sierra AI serves 40% of the Fortune 50 with outcome-based pricing and no public rates, with year-one costs reportedly reaching $200K-$350K+. Outcome pricing at enterprise scale works when volume smooths the variance. For startups without that volume, a single high-cost outlier can erase a month's margin.
Hybrid: Subscription Base With Usage Components
Most AI products land here. A base subscription provides revenue predictability while usage components capture value as customers scale.
GitHub Copilot charges $10/month Individual, $19/user/month Business, and $39/user/month Enterprise. Microsoft Copilot for Security bills $4/hour of compute. Notion bundles AI into its $20/user/month Business tier.
Three tiers cover most products: a starter tier with low price and limited usage, a growth tier with higher limits that's the obvious choice for most customers, and an enterprise tier priced on negotiation. Hide pricing complexity for 90% of users. Let heavy users buy additional capacity after hitting their plan limit, rather than making everyone a meticulous gas-meter tracker.
Match the Model to Your Product's Maturity
Before committing, be honest about where the product sits. Commodity products compete on price, with pure usage at the lowest unit cost; most wrapper and thin-layer AI products live here whether they admit it or not. Differentiated products compete on quality, with premium per-unit pricing justified by measurable quality differences. Indispensable products compete on outcomes and can make delivery promises because their output variance is narrow enough to keep them.
Most companies believe they're differentiated. Most are commodity. Commodity products doing outcome pricing make promises they can't keep, and indispensable products on pure usage pricing leave money on the table. If you don't know which level you occupy, you probably lack the cost and outcome data to tell. That is itself the answer: start with usage-based or hybrid pricing and measure your way toward outcome pricing as your output variance narrows.
How High Can the Price Go?
Cost data is the floor. Value is the target. The ceiling is set by alternatives, and that ceiling is moving.
If your AI replaces a $50,000/year employee, $10,000 annually is a strong value proposition regardless of delivery cost. If you prevent a $100,000 compliance violation, $10,000 is rational risk reduction. A useful rule of thumb: aim for 5-10x ROI for the customer. At 10x the purchase decision is obvious. Below 3x it becomes a negotiation.
Customers compare your price to four alternatives: hiring someone to do the work manually, using a competitor, building it themselves, or doing nothing.
That third option is changing faster than most vendors have noticed. Retool's 2026 report found that 35% of their customers have already replaced at least one SaaS tool with a custom build, and 78% expect to build more of their own tools in 2026.
The practical test: could your customer's team member, armed with Cursor and a weekend, replicate 80% of your core workflow? When that build option was $500K and 12 months, paying $50K/year for SaaS was obvious. When it costs $5,000-$15,000 and three months with AI-assisted development, the calculus changes.
This doesn't mean underprice. Price signals quality, and in B2B a $29/month AI product competing against $500/month solutions raises suspicion rather than excitement. But the ceiling on what you can charge is a moving target, and it's moving down. Anchor pricing to alternatives, not to what the market would theoretically pay.
Free Tiers Come With an Invoice
Free trials and freemium reduce the first purchase decision to zero. For AI products, that comes with a specific catch: you pay real, per-interaction costs for users who haven't committed to paying.
Free trials grant full access for 14-30 days and put a clock on the decision. They work best when users experience value within minutes. One design choice helps a lot: start the trial clock at the user's first meaningful action, not at account creation.
Freemium works when usage reveals value over time and the cost to serve free users stays low. For AI products, "low cost to serve" is the binding constraint, so set clear boundaries: limit AI interactions per month, restrict which models or features are available, cap compute time or output length.
Monitor cost to serve free users and free-to-paid conversion together. If you spend $5/month per free user with a 2% monthly conversion rate, you're spending $250 in free-tier costs per converted customer. That's your effective CAC from the free tier. If paid ads acquire customers at $150 and the free tier acquires them at $250 with higher engagement, the free tier may still win on LTV. If it acquires at $250 with lower engagement, it's a cash furnace dressed up as a growth engine.
Signals the Price Is Wrong
Pricing isn't a launch decision. It's a continuous process. In the first year, revisit quarterly. As the product matures, semi-annual reviews are sufficient unless model costs shift sharply, which they often do.
Four signals suggest your price is too low: customers accept without negotiation, sales cycles are unusually fast, buyers describe it as "a steal," and churn is high among low-engagement users who never valued what they barely paid for.
Four signals suggest it's too high: frequent discount requests, high abandonment at the payment step, sales cycles stuck at procurement, and declining win rates despite strong product fit.
When adjusting, raise prices for new customers first and measure conversion impact before applying broadly. Grandfather existing customers or give 90+ days notice with clear communication about added value. Test packaging changes before testing price changes, because packaging often moves the needle more than the number itself.
What Changes Once You Have the Data
A product manager is deciding what belongs in the $49 starter tier. Per-feature cost data shows chat runs about $0.003 per request while document analysis runs $0.12. Document analysis moves to the growth tier, and the starter tier stops depending on light usage to stay profitable.
An engineer restructures a retrieval workflow so the shared system prompt hits the provider's cache. Cached input bills at a 90% discount, and cost per trace drops by a third with no quality change.
A founder preparing a board deck breaks gross margin down by plan tier instead of reporting the blended average. The $49 tier runs at 22% margin, the $199 tier at 71%. The next pricing conversation stops being about the average and starts being about which tier to repackage.
None of these people touched a price list. All three moved the economics under one, because AI decisions are pricing decisions.
Every Pricing Argument Traces Back to the Same Gap
The AI pricing discourse is full of confident, mutually exclusive claims. Seat-based pricing is dead; seat-based pricing improved retention. Outcome-based pricing is the future; outcome-based pricing is a buzzword. Usage-based pricing is necessary; usage-based pricing tanks retention. Credits are a growth lever; credits are customer-hostile.
Lay those arguments side by side and the contradictions aren't subtle. Almost every disagreement traces back to the same gap: per-request, per-customer cost-to-serve data barely exists yet, so the arguments run on intuition, anecdotes, and small samples.
Truthfully, the market may be converging on hybrid pricing not because it's the best model, but because it's the model you can implement without knowing your true cost-to-serve. Companies that do have per-request cost data might arrive at different answers entirely: flat pricing for segments that are reliably profitable, usage gating for the ones that aren't, outcome pricing for the workflows where output variance is narrow enough to support it.
If You Only Do One Thing This Week
Skip the framework and build the dataset. Export revenue by customer from Stripe. Pull provider usage by API key from your OpenAI and Anthropic dashboards. Join the two in a spreadsheet and compute cost as a percentage of revenue, per customer. The first pass takes about four hours and it won't be precise. It will still be your first honest look at which customers make money. Every decision in this guide gets easier once that column exists. Bear Lumen automates the same join, cost-to-serve per customer, per feature, and per outcome, if you'd rather not maintain it by hand.
Cost-to-serve data is what turns pricing from guesswork into architecture. The teams that measure it get to design their pricing. The teams that don't get to react to it.