Pricing an AI Feature: Per-Seat vs. Per-Use vs. Credits

AI features cost real money per request. If you price like classic SaaS ($20/seat/month unlimited) and one power user runs a 10M-token job, you just lost money on them.

• Per-seat: great when usage is predictable and bounded (daily summaries)
• Per-use / metered: best when usage is spiky or power-law (video generation)
• Credits: hybrid — users feel in control, you protect margin with expiry
• Freemium with hard rate limits: use only if free users drive real virality
• Always model gross margin at 60%+ AFTER inference cost, not before

Watch your 90th-percentile user. If their inference cost exceeds your plan price, either raise the ceiling, switch to credits, or cap the feature.

# Simple margin math Plan price: $29/user/month Avg inference cost: $4/user/month P90 inference cost: $22/user/month <-- danger zone Stripe + infra: $3/user/month Gross margin (avg): ($29 - $4 - $3) / $29 = 76% OK Gross margin (P90): ($29 - $22 - $3) / $29 = 14% NOT OK Fix: add a 500-credit cap, overage at $0.05/credit.

You're pricing right when 80%+ of customers feel the plan is fair, your P90 user still pays more than they cost, and you don't panic when Anthropic has a latency spike.