Price an AI feature from the value it delivers. Set the floor from what it costs to serve. Measure AI cost per unit twice, once for the typical user and once for your heaviest cohort, and divide by one minus your target AI margin to get the floor. Then include an allowance sized for typical use and meter everything above it with credits or overage priced at or above that floor. In Kyle Poyar’s 2026 Growth Unhinged survey of 230 software companies, internal costs and margins were the top factor in pricing AI capabilities (54%), and the median target AI margin was about 50%. If you have not picked the charge metric yet, start with seat vs usage vs outcome pricing. This page covers the next step: making the price cover the inference bill.
Pick the package before you set the number
| Package | Fits when | Who carries heavy-user cost | Public example |
|---|---|---|---|
| Bundle into a higher tier | AI use is light and even; you want adoption | You, inside a flat price | Notion: core AI in Business at $20 per member a month since May 2025; Free and Plus get a trial |
| Paid add-on per seat | Procurement buys seats and value tracks the person | You, unless the add-on has a cap | Microsoft Copilot add-on about $30 on top of an Office 365 license (Bessemer) |
| Included allowance plus overage | Usage per user is uneven | The heavy user, above the allowance | GitHub Copilot Pro: $10 includes $10 of AI credits plus a flex allotment; set a budget for more |
| Prepaid credit packs | Agents and automations run with no person in the loop | The buyer, by pack | Salesforce Flex Credits: $500 per 100,000; one action is 20 credits ($0.10). Notion Custom Agents: $10 per 1,000 credits |
| Per conversation or outcome | Success is unambiguous and you can absorb variance per task | You, per task | Salesforce Agentforce $2 per conversation; Intercom Fin $0.99 per resolution (Bessemer) |
Bessemer’s AI pricing playbook names usage caps, prepaid credits, hybrid models, and tiered subscriptions as the tools for balancing variable compute against predictable revenue. Poyar’s fixes for weak expansion point the same way: a premium edition priced 50–100% above the base, add-ons that some customers value, and consumption limits that hit the top 10% of power users. The last row is a charge-metric decision. Work it on seat vs usage vs outcome pricing.
Work out the cost floor per unit
Pick the unit you charge for: a seat, a task, a credit, or a conversation. Count the AI delivery costs that belong in COGS: production inference, production GPU, retrieval and vector serving, monitoring, AI support, and human review of customer output. Leave dev/test inference and training out. The classification table lives on AI SaaS gross margin, and the AI gross margin calculator returns cost per user and the heavy-cohort share from your monthly numbers.
Then apply the formula to both costs:
| Target AI margin | Floor as a multiple of AI cost per unit |
|---|---|
| 40% | 1.67× |
| 50% | 2× |
| 60% | 2.5× |
| 70% | 3.33× |
Pick the target before you see the result. Poyar’s median target is about 50%. 12% of his sample aim for SaaS-like AI margins of 80%+, and a similar share aim for 20% or lower (he names PostHog). Bessemer puts AI gross margins at 50–60% against 80–90% for SaaS. Murray’s published bands by product type are on the gross margin page.
The floor is the minimum price. Bessemer says most founders default to cost-plus (calculate costs, double it) and undercharge because asking for more feels awkward. Anchor the price against what the problem costs the buyer today. Keep the cost per unit of work in view so an outcome or flat price cannot invert your margin at scale (FounderNexus session). There is no reliable public benchmark yet for the markup AI companies charge over inference cost on credits or overage.
A labeled example: flat bundle vs allowance plus overage
A B2B product adds an AI assistant to 1,000 paid seats. Measured AI COGS per month: 900 typical users at $2.50 each and 100 heavy users at $60 each, $8,250 in total. The heavy cohort is 73% of the bill, close to Poyar’s 70%+ observation. The target AI margin is 50%.
| Line | Flat bundle | Allowance plus overage |
|---|---|---|
| Price design | $16.50 seat uplift (average cost × 2) | $10 seat uplift with $5 of AI usage included; overage at 2× cost |
| Typical user: price vs AI cost | $16.50 vs $2.50 | $10 vs $2.50 |
| Heavy user: price vs AI cost | $16.50 vs $60 | $120 ($10 + $55 overage × 2) vs $60 |
| AI revenue vs AI COGS | $16,500 vs $8,250 | $21,000 vs $8,250 |
| AI margin | 50% | 61% |
| AI margin if heavy-user cost doubles | 14% | 57% |
The flat bundle hits the target on day one and loses $43.50 a month on each heavy user. If agentic features double heavy-user cost, its margin falls from 50% to 14% and the price has not moved. The allowance plan charges typical users less, keeps each heavy user at a 50% margin, and holds the blend near target when usage climbs. The doubling row is a stress test. GitHub and Cursor both changed pricing after agentic and long-horizon use pushed per-request costs up.
Size the allowance and price the overage
Size the allowance from median use. Cursor describes its Pro allowance in median terms. Its July 2025 post says that at median token usage $20 covered about 225 Sonnet 4 requests, 550 Gemini requests, or 650 GPT 4.1 requests, and that the vast majority of Pro users do not run out. GitHub sets base credits equal to the subscription price and adds a flex allotment that it says will adapt as model pricing and efficiency change. Unused credits expire at month end. Business plans pool them across the organization so no seat strands capacity.
Price the overage above the floor. GitHub bills extra usage at published API rates and Cursor sells more at cost. Both earn their margin on the plan. A startup with thin plan margin cannot copy that. Price overage at the floor or above, in blocks the buyer can approve once.
Give each credit a fixed dollar value. GitHub fixes one AI credit at $0.01. Salesforce fixes one Agentforce action at 20 credits. Poyar warns that credits are great for vendors and can become a nightmare for customers who manage different credit models across dozens of vendors. A credit the finance team can convert to dollars survives procurement.
Report overage apart from the subscription. Usage above a commitment is not ARR. Keep it on its own line (usage revenue as ARR).
Guard the margin on heavy users and agents
Give admins a budget. GitHub lets admins set budgets at the enterprise, cost center, and user level, then choose whether to allow more usage at published rates or cap spend. Notion pauses Custom Agents when credits run out until an admin adds more.
Price agent runs apart from human seats. Notion keeps core AI inside the $20 Business seat and meters Custom Agents in credits from 4 May 2026. GitHub’s stated reason for its switch: a quick chat question and a multi-hour autonomous coding session could cost the user the same amount, and the old request model was no longer sustainable. Cursor said its hardest requests cost an order of magnitude more than simple ones.
Route work to cheaper models. ICONIQ’s January 2026 survey, via SaaStr, found model inference rising from 20% to 23% of spend as products scale, and the best teams sending most tasks to smaller models. GitHub gives paid users a 10% discount on model costs when they use auto model selection.
Reprice without losing trust
| Company | What changed | How they handled existing customers |
|---|---|---|
| Cursor | 16 Jun 2025: Pro moved from 500 requests to $20 of usage at API pricing | 4 Jul 2025 apology; refunds for unexpected usage from 16 Jun to 4 Jul |
| GitHub | Announced 27 Apr 2026; usage-based billing live 1 Jun 2026 | Base prices unchanged; preview bill in early May; Business $30 and Enterprise $70 in promotional credits for June to August |
| Notion | 13 May 2025: AI bundled into Business and Enterprise | New prices for new subscriptions only; existing customers moved at first renewal after 13 Aug 2025 |
Expect to do this more than once. Poyar found that three in four software companies changed pricing or packaging in the past year.
Build the AI price in order
AI feature price build
- Measure AI COGS per unit. Use the COGS lines on the gross margin page. Split typical users from the heavy cohort.
- Write the target AI margin down. One number, agreed with whoever owns the board pack.
- Compute the floor. Cost per unit ÷ (1 − target). Do it for the typical and the heavy unit.
- Anchor to buyer value. Price against what the problem costs the buyer today. Use the floor as a check on that number.
- Pick the package. Bundle, add-on, allowance plus overage, credit packs, or per outcome, from the table above.
- Size the allowance at typical use. Median user plus headroom. Check what share of users would exceed it.
- Price overage or credits at or above the floor. Fix a dollar value per credit and sell in blocks.
- Add budgets, caps, and alerts. Admin budgets, spend caps, and a usage warning before the cap.
- Announce with notice. Preview bills, a date, and current customers moved at renewal.
- Report it monthly. AI margin, heavy-cohort share, and overage revenue as their own lines. Reprice when AI cost per unit or tokens per task moves.
Mistakes that cost margin or trust
Shipping AI free inside the plan with no cap. The heavy cohort sets your cost, and a flat price does not move with it.
Setting the price at cost × 2. That is the floor at a 50% target. Bessemer calls cost-plus a trap.
Pricing from average cost. The example above hits 50% on average and loses money on each heavy user.
Selling overage at cost with no plan margin. GitHub and Cursor can do it because the plan carries margin. Check that yours does.
Credits with no dollar anchor. Buyers cannot budget them and procurement slows the deal.
Switching overnight. Cursor’s 2025 change ended in an apology and refunds.
Counting overage as ARR. Report it as usage revenue next to committed ARR.
This page is operator judgment on pricing design, not legal, tax, or accounting advice.
Sources
- Growth Unhinged, The 2026 State of B2B SaaS and AI Monetization Report. Kyle Poyar, 13 May 2026. 230 B2B software and AI companies. Internal costs and margins the top factor in pricing AI (54%); market 36%; productivity gains 30%. Median target AI margin about 50%; 12% aim for 80%+; a similar share 20% or lower (PostHog). AI credits 29% today; 33% plan to add within 6 to 12 months; credit models grew 126% year on year in 2025. Premium edition 50–100% above base; consumption limits for the top 10% of power users (top 10% drive 70%+ of token consumption, in Poyar’s experience). Three in four companies changed pricing or packaging in the past year. Hybrid 37%.
- Growth Unhinged, Why everyone’s switching to AI credits. Kyle Poyar, 3 September 2025. Credit models across Microsoft, Salesforce, Cursor, and OpenAI; credits let vendors include some AI use free and adapt to model costs.
- Bessemer Venture Partners, The AI pricing and monetization playbook. Atlas Editors, 10 February 2026. Pricing must account for inference costs; 50–60% AI gross margins vs 80–90% SaaS; cost-plus (calculate costs, double it) undercharges; platform fee at 2X delivery costs plus outcome credits; usage caps, prepaid credits, hybrid models, and tiered subscriptions to manage variable compute; Intercom Fin $0.99 per resolution.
- Bessemer Venture Partners, Part III: Business model invention in the AI era. 5 December 2024. Microsoft Copilot add-on about $30 on top of Office 365 licenses.
- The SaaS CFO, What Should Be Included in AI COGS. Ben Murray, posted 5 May 2026, updated 25 September 2026. AI COGS categories; heavy users 50 to 100 times light users (citing Agicent and SaaStr).
- GitHub Blog, GitHub Copilot is moving to usage-based billing. Mario Rodriguez, 27 April 2026. Switch on 1 June 2026; credits at published API rates; Pro $10 with $10 of credits; Business $19 and Enterprise $39 per user unchanged; promotional $30 and $70 for June to August; preview bill in early May; pooled usage; admin budgets; quick chat and multi-hour session priced the same under the old model.
- GitHub Docs, Usage-based billing for individuals. Base credits plus flex allotment; 1 AI credit = $0.01; unused credits expire monthly; 10% discount with auto model selection.
- Cursor, Clarifying our pricing. Michael Truell, 4 July 2025. Pro: $20 of frontier usage at API pricing, more at cost; previously 500 requests; hardest requests an order of magnitude costlier; about 225 / 550 / 650 requests at median usage; refunds for 16 June to 4 July.
- Notion, Introducing Notion AI for Work. John Hurley, 13 May 2025. AI bundled into Business at $20 per user per month; existing customers move at first renewal after 13 August 2025.
- Notion, Pricing and Notion AI. Checked 16 October 2026. Custom Agents $10 per 1,000 Notion credits; credits apply from 4 May 2026; agents pause when credits run out.
- Salesforce, Salesforce Introduces New Flexible Agentforce Pricing. 15 May 2025. Flex Credits $500 per 100,000; one action 20 credits ($0.10); conversations $2 each.
- SaaStr, 5 Key Takeaways from ICONIQ’s State of AI Report. Jason Lemkin, 7 February 2026. ICONIQ January 2026, about 300 executives; inference 20% to 23% of spend as products scale; route most tasks to smaller models.
- FounderNexus. Founder sessions, 2026. Anchor price to the buyer’s current cost of the problem; know variable cost per unit of work so outcome pricing does not invert margin. Not a survey. No closed-session numbers.
Related
- Go-to-market
- Seat vs usage vs outcome pricing for AI SaaS
- AI SaaS gross margin: inference COGS in the board pack
- AI gross margin calculator
- Does usage revenue count as ARR?
- How to convert a pilot to a paid contract
Founders who have repriced an AI feature and kept their customers will pressure-test your allowance and overage math in a FounderNexus session.