How to Price AI Features in Your SaaS (Without Losing Margin)

Every SaaS founder shipping their first AI feature runs into the same question: how do I price this without watching my margin evaporate? Here are six pricing patterns that work in 2026, when to use each, and the exact margin math to protect the business.

The margin problem

Traditional SaaS margins run 75–85%. Naive AI SaaS often runs 30–50% because a $20/mo customer can burn $18 of API calls in a bad month. If you don't price for it, one power user takes down your unit economics.

Six pricing patterns

1. Feature-gated tiers

AI features live only in higher plans. Simple, familiar, but leaves money on the table.

• Best for: teams new to AI, unclear usage patterns.

2. Credit / token buckets

Customers get X credits per plan, buy more if they run out.

• Best for: image gen, transcription, anything with clear "one action = one credit" mapping.

3. Message / action metering

Charge per AI interaction (chat message, generation, agent run).

• Best for: chatbots, agents, workflows where each action has consistent cost.

4. Hybrid: seat + usage

Flat seat price includes a fair-use allowance. Overage is metered.

• Best for: B2B tools with mixed light and heavy users.

5. Model-tier pricing

Basic plan uses cheap models, higher plan unlocks premium models.

• Best for: use cases where quality genuinely varies by model (coding assistants, deep research).

6. Outcome-based pricing

Charge per meeting booked, ticket deflected, lead qualified.

• Best for: mature products where outcome attribution is clean and defensible.

The margin math

Target: 65–80% gross margin on the AI line item after Year 1.

1. Measure your average cost per event (token spend + infra + storage).
2. Set the retail price at 3–5× cost.
3. Cap heavy users with fair-use limits or overage pricing.
4. Re-run the math every quarter — API prices move.

Example: a support bot costs $0.02/message all-in. Fair pricing is $0.06–$0.10/message wholesale, packaged as "500 messages/mo included in $99 plan, $0.05/message after."

How to structure overage

• Soft cap → email at 80%, warn at 100%, auto-upgrade or throttle at 110%.
• Hard cap → hard stop unless customer opts into overage billing.
• Never let a $99/mo user drive $1,000 in API costs unattended.

Cost controls that protect margin

• Cache aggressively — cuts spend 40–70% on repetitive workflows.
• Route simple requests to cheap models — see provider comparison.
• Batch when you can — providers often discount async batch APIs by 40–50%.
• Set per-user daily limits by default.
• Monitor per-customer P95 spend weekly and investigate outliers.

What not to do

• Don't hide AI pricing. Users hate surprise bills more than transparent metering.
• Don't set the same cap for all plans. Enterprise users need real headroom.
• Don't price 20× your cost. Buyers benchmark against ChatGPT ($20/mo) — a 100× markup gets ignored.
• Don't switch models silently. If quality changes, your customers will feel it.

Communicating the price

Two rules: (1) show value per unit, not tokens or credits — customers understand "500 support conversations" better than "5M tokens"; (2) always show the fair-use number in the plan card, not the fine print.

Grandfathering

When you re-price (and you will), grandfather existing customers for 6–12 months on the old plan. Cheaper than the churn. Announce well in advance.

The forever rule

Your AI margin should hold at 65%+ even if API prices double overnight. If that math breaks, your fair-use caps are too generous or your prices are too low. Fix both before you scale marketing spend.