In the book I told the story of Close.com, a CRM that once included unlimited free phone calls in every plan. Lovely for customers. Less lovely for Close, because every call cost it money. Eventually it pulled calls out of the plan and charged per call.
With AI, every feature is a phone call.
And the problem isn't your average customer. It's the 10% who use your product ten times as hard. This calculator shows you their margin, not just the average.
How to use the AI gross margin calculator
- Describe one task. Choose the model, the input and output tokens for a typical request, and any share served from cache.
- Describe your users. Set the tasks a median user runs each month, then the share of light, median and heavy users and how much more (or less) each uses. If you have usage data, use your 25th, 50th and 95th percentiles.
- Add other costs. Hosting, vector database, third-party APIs and support per user. Payment fees default to 2.9% plus 30c per charge, so edit them to your provider's rate.
- Choose how you charge. A flat plan, a plan with an allowance and overage, or credit packs that heavy users buy more of.
- Stress-test. Put 100 in "model price change" to see a price doubling, then try routing tasks to a cheaper model.
How the maths works
Cost of goods sold (COGS) includes everything it takes to serve a customer: inference, hosting, APIs, support and payment fees.
Cost per task = [Input × (1 − cache) × input price + Input × cache × cache price
+ Output × output price] ÷ 1,000,000
Segment COGS = Tasks × cost per task + other costs + payment fees
Gross margin = (Revenue ex GST − COGS) ÷ Revenue ex GST
Break-even price = (Average COGS before fees + fixed fee) ÷ (1 − fee %)
The blended margin weights each segment by its share of users. If you tick GST, the tool strips the 10% out of revenue first, because GST you collect isn't yours.
A worked example
The pre-filled example is the $49 AI writing assistant from our AI gross margin guide. A median user runs 600 tasks a month on Gemini 4 Argon, at 4,000 tokens in and 700 out. That's $0.015 a task, or $9 of inference. Add $4.30 of hosting, APIs and support and $1.72 of card fees, and COGS is $15.02. Margin: 69.3%.
Healthy. Until you look at the heavy row.
A heavy user on five times the median runs 3,000 tasks. Their inference alone is $45, so they cost about $51 a month on a $49 plan. The light users are paying for them. Now type 100 in the price change box (Google's pricing page says Argon doubles on 1 December 2026) and watch the blended margin drop by about 20 points.
Then route 50% of tasks to Claude Haiku 5.5. The margin comes straight back, even at the higher price.
How to read the result
The guide cites these 2025–26 benchmarks: traditional SaaS at 75–80% gross margin and AI-native companies at 50–60% (SaaStr, February 2026), with the average AI product at 45% in 2025, projected 53% for 2026 (ICONIQ, via SaaStr). Bessemer put its fastest-growing AI companies at about 25%.
So a blended margin above 60% is good for AI. But the segment rows matter more. A negative heavy-user row is a pricing problem, not a cost problem. The fix is usually an allowance with fair overage, or credits, so that your best customers pay for what they use.
The flat-plan "task ceiling" in the verdict is the number of tasks at which one user stops paying for themselves. Put your fair-use limit below it.
What to do next
- Turn your cost per task into a rate card with the Credit Pack Designer.
- Compare the same workload across 20 models with the AI Token Cost Calculator.
- Read how to price an AI wrapper if you resell one vendor's model.
- See the full picture in AI pricing: the complete 2026 guide.
- Gross margin drives what your business is worth. See SaaS valuation multiples. In Australia, you can talk through what recurring revenue could mean for your business in a free Subscription Profit Review.
Frequently asked questions
What is a good gross margin for an AI company?
Benchmarks cited on this site put AI-native companies at roughly 50–60% gross margin, against 75–80% for traditional SaaS (SaaStr, February 2026). ICONIQ's survey put the average AI product at 45% in 2025, projected to reach 53% in 2026. Above 60% with all inference in COGS is a strong result for an AI product today.
Should inference costs go in COGS?
Yes, the inference you run to serve paying customers belongs in cost of goods sold, alongside hosting, third-party APIs, support and payment fees. Training and R&D experiments usually sit in operating expenses. Leaving inference out flatters your margin, and any serious buyer or investor will put it back when they assess the business.
How do I find my heavy users' usage?
Export tasks or tokens per customer for the last month and sort them. The median is your middle customer; the heavy segment is roughly the top 5–10%. If you're pre-launch, assume heavy users run five to ten times the median. AI products often see a long tail of power users and automation that dwarfs typical use.
Do credit packs fix low AI margins?
They help, because revenue rises with usage and heavy users pay for what they consume. But credits only protect margin if each credit is priced above its cost, including on your most expensive action at your cheapest bulk rate. Switch the calculator to credit packs and check the heavy-user row before you launch.
Why does GST change my margin?
In Australia, if your price includes GST, one-eleventh of it belongs to the ATO, not to you. Gross margin should be calculated on the GST-exclusive amount. Card fees, though, are charged on the full amount the customer pays, so they take a slightly bigger bite than the headline rate suggests.
Built on The Subscription Playbook
This tool turns frameworks from Robert Coorey's book into something you can use in minutes. Robert is the author of The Subscription Playbook and a Director at BPAG Accountants. More about Robert.