An Honest Method for Pricing AI Features in 2026

Pricing AI features gets done badly because the methods imported from traditional SaaS don't account for the structural variance in token consumption. A customer paying $29/month for an AI tool might cost $2 in tokens or $40, depending on usage. Traditional SaaS pricing methods (competitor benchmarking, value-based, willingness-to-pay surveys) all assume roughly uniform cost-per-customer, which AI features structurally violate. This article lays out a method designed for that violation.

1. The six-step method

The method, in order:

1. Measure token-cost variance across customers. The single most important input. Most founders skip this and pay for it in pricing-model misfit.
2. Segment customers by intent and willingness-to-pay. Light/medium/heavy usage maps to different pricing strategies. Conflating them produces a pricing model that fits none of them.
3. Pick the pricing model. Cost-plus, value-based, or hybrid. The 40% variance rule decides.
4. Set the price band. Floor (margin target), ceiling (competitor band), middle (the launch price). Each calculated, not guessed.
5. Test with a holdout group. A/B the new pricing on a fraction of new sign-ups. Measure conversion, churn, and ARPU separately.
6. Iterate quarterly. Token cost shifts, customer mix shifts, competitor pricing shifts. Re-run quarterly minimum.

The method takes 4-8 weeks to execute end-to-end the first time, and 1-2 weeks for each quarterly re-run. The investment pays back in the first pricing change that fits the data rather than the founder's intuition.

2. Step 1: Measure token-cost variance

Pull the last 30 days of API logs per customer. Compute monthly token cost per customer. Calculate:

• Median token cost per customer. This is the typical cost — what most customers consume.
• 90th percentile cost. What the top 10% consume.
• Variance ratio. 90th percentile divided by median. This is the variance metric the method depends on.

If variance ratio is below 2x, flat-monthly pricing is appropriate. Between 2x and 4x, consider tiered pricing (3 fixed tiers). Above 4x (the typical AI-feature case), hybrid pricing becomes the structurally right answer. NYU Stern's software-industry margin data^[4] shows that products with high cost-variance and uniform pricing have systematically lower gross margins than products that match pricing model to variance.

Most AI products land at variance ratios of 5-15x. The Anthropic system-prompt sample pattern (some users hammer the API with continuous queries; most don't) produces this distribution. Heavy users justify usage-pricing; light users get priced out by usage-pricing. Hybrid resolves both.

3. Step 2: Segment customers by intent

Segmentation should be based on observed behavior, not survey responses. Plot customers on a 2x2: vertical axis token usage (low/high), horizontal axis willingness-to-pay (low/high). Four cells:

• Low usage, low WTP: The casual users who churn fast. Don't price for them; let them in on a free tier or low-cost monthly plan.
• Low usage, high WTP: The professionals who use sparingly but pay for reliability. Flat-monthly works perfectly.
• High usage, low WTP: The price-sensitive power users. Usage-based pricing prices them out; flat-monthly subsidizes their usage with other customers' margins.
• High usage, high WTP: The ideal customers. Hybrid pricing captures the most revenue from them without breaking the flat-monthly tier.

The segment distribution informs pricing model selection. If the customer base is dominated by the low-WTP power-user cell, the pricing problem is structural — the product is attracting price-sensitive heavy users and there's no pricing model that solves it without losing them. The right move is product-level positioning to attract a different segment, not pricing tweaks.

4. Step 3: Pick model — the 40% rule

The 40% rule: if the standard deviation of token cost per customer divided by the mean token cost is above 40%, hybrid pricing dominates. Below 40%, flat-monthly or simple tiered pricing wins.

Why 40%? Below that variance, the per-customer cost is predictable enough that flat-monthly captures most of the available margin without the operational complexity of usage tracking. Above 40%, the highest-spending 20% of customers consume 40-60% of total token cost — flat-monthly either underprices them (founder loses margin) or overprices everyone else (founder loses conversion). HBR's pricing research has documented similar variance-to-pricing-model mapping in other industries^[2]; MIT Sloan reports the same in industrial pricing^[3].

The three model options:

• Cost-plus: Per-token markup with a small base subscription. Best when token cost is the dominant value driver and customers are technical enough to understand it. Common in developer tools.
• Value-based: Flat-monthly priced against the customer's outcome (revenue lift, time saved). Best when token cost is small relative to the product's economic value. Common in B2B SaaS.
• Hybrid: Base subscription + usage overage. Best when the customer mix spans light and heavy users. Default for most AI products.

5. Step 4: Set the price band

Three price points to compute:

• Floor: Cost / (1 − target margin). At $4 per-user cost and 75% target margin, floor is $16/month.
• Ceiling: 2x the highest competitor in the comparable band. Ceiling pricing is rarely the launch price but worth knowing as the upper bound on willingness-to-pay.
• Launch: Between floor and median competitor price. Bessemer's 2024 cloud benchmarks place healthy SaaS pricing at 1.5-2.5x cost^[1]. At $4 cost, that suggests $6-$10 floor and $20-$30 launch.

For hybrid models, set the base subscription at roughly the cost floor for the median user, then price overage at 1.5-2x marginal cost. The base captures the predictable revenue; the overage captures the heavy-user variance.

6. Step 5: Test, measure, adjust

The right test design:

• A/B the pricing on new sign-ups only. Grandfather existing customers. Run for 90 days minimum to capture both conversion and early-churn signal.
• Measure three metrics separately. Conversion rate (do they buy), ARPU (what do they pay), 90-day retention (do they stick). Pricing changes can lift one and hurt another; the net is what matters.
• Pre-commit the success threshold. "Adopt the new pricing if 90-day blended revenue clears X% of the control." Define X before running the test.

ChartMogul's retention data shows that pricing changes typically have a 4-6 month tail before retention impact is fully visible^[5]. Don't make permanent decisions on 30-day data.

7. Step 6: Iterate quarterly

Three inputs drift quarterly:

• Token cost. Model price changes (drops or shocks), prompt-caching adoption, routing optimizations, output cap changes. All move the per-customer cost over time.
• Customer mix. Acquisition channel shifts can change the underlying willingness-to-pay distribution. A founder who scales paid ads will attract a different mix than one who scales content.
• Competitor pricing. Vendor and substitute pricing moves. The ceiling shifts.

The quarterly re-run is short — 1-2 weeks of analysis followed by a small adjustment to the price band or model. The cumulative effect over 4-8 quarters is a pricing strategy that fits the current reality rather than the founder's day-one intuition.

One operational rule. Pricing changes should be small and frequent (5-15% adjustments quarterly) rather than large and rare (50% adjustments annually). Small frequent changes preserve customer trust and produce continuous learning; large infrequent changes produce churn shocks and reactive customer reactions. The method's quarterly cadence is designed to enable small frequent adjustments.

One additional pattern worth pre-empting: founders who skip the variance measurement (step 1) and jump straight to model selection (step 3) consistently pick the wrong model. The structural data is what the decision needs to anchor on. Skipping it produces pricing choices that feel reasonable but fail under the real customer distribution. Run the variance calculation even if it takes a day of log-parsing — the rest of the method is wasted without it.

An extended pattern worth pricing: the relationship between pricing and product roadmap. Pricing changes that aren't backed by product changes get reversed within 6 months. If the founder raises prices from $29 to $39 without shipping additional features or improving the perceived value, the new price erodes through retention churn and conversion drops. The honest pattern is to bundle price changes with product changes: raise the price when shipping new capabilities, lower the price (or hold it) during pure-maintenance quarters. This ties pricing to the product strategy rather than treating it as an independent lever.

A second extended pattern: the right pricing for AI features is rarely the same as the right pricing for non-AI SaaS features in the same product. A founder running an AI-heavy product with both AI features (heavy token usage) and traditional CRUD features (zero marginal cost) often prices them at the same monthly tier, which structurally underprices the AI features and overprices the non-AI features. The cleaner approach is to separate them: an AI feature add-on (priced at variance-aware hybrid) on top of a core product subscription (priced at flat-monthly). The complexity is worth it when the AI features represent more than 30% of token-driven cost.

A third pattern about communication. Pricing changes that customers don't understand drive churn. A flat-monthly to hybrid migration that customers experience as "now they're charging me usage on top of the subscription" feels like a price increase even when the math works out the same or better. The communication is the product. Run customer interviews on the new pricing before launching it. If users can't articulate why the new pricing is fair, the pricing strategy is incomplete regardless of how well it fits the variance math.

One operational caveat about the 40% variance rule: the threshold holds up at typical AI-product scales (1,000-50,000 active users) but breaks down at very small samples. A product with 50 active users that shows 60% variance is more likely showing random sampling noise than a real cost-distribution signal. The variance threshold needs at least 200-500 customer-months of data to be reliable; below that, default to flat-monthly and revisit when the data is enough.

The Pricing Model Picker handles the model-selection step computationally. The AI Product Margin calculator handles the cost-side input. The SaaS Pricing Strategy calculator handles the price-band step. Together they implement the quantitative pieces of the six-step method; the founder still owns the segmentation and iteration steps.

Frequently asked questions

When does hybrid pricing win for AI features?

When token-cost variance per customer exceeds 40%. Below that, flat-monthly is structurally simpler and converts better. Above 40% variance, hybrid (base subscription + usage overage) is the only model that doesn't leave money on the table from heavy users or price out light users.

How is this method different from generic SaaS pricing advice?

It treats token cost as a real variable, not a rounding error. Generic SaaS pricing assumes uniform cost-per-customer; AI products have 10-50x variance in per-user token cost between light and heavy users. The method explicitly measures that variance and chooses the pricing model that fits.

Why six steps instead of three or ten?

Three steps oversimplify the customer-segment work; ten steps add ceremony without signal. Six steps cover the structural decisions (cost measurement, segmentation, model choice, price band, testing, iteration) without padding with low-impact activities like 'create a pricing committee' that don't apply at solo scale.