Unit Economics

The Unit Economics of AI Features: Is Your Chatbot Actually Profitable?

Cost per request, cost per user, margin per customer: how to work out whether an AI feature makes money, why the heavy users decide the answer, and a calculator to run your own numbers.

5 minBy LLMtrackUpdated Rates checked 17 Nov 2026
Monthly LLM cost of a heavy user making 3,000 chat requests on a $29 plan: $0.81 on GPT-6 Luna, $8.10 on GPT-6.1 Terra, $16.20 on Claude Sonnet 5.5 and $32.40 on Claude Opus 5.5, which is more than the plan price.
Short answer: LLM cost per user = cost per request × requests per user per month. Compare it with the plan price for a typical user and for a heavy one, not just the average. In the worked example, a $29 plan on GPT-6.1 Terra costs $0.81 a month for a typical user and $8.10 for a heavy one, so margins after LLM cost are 97% and 72%. The same heavy user on Claude Opus 5.5 costs $32.40 and loses money. Measure cost per customer from real requests, then decide on caps, model tiers or pricing.

A subscription product with an AI feature has a cost that grows with use and a price that doesn't. The more a customer uses the feature, the more they cost you, while they keep paying the same plan price. For most customers that's fine. For the heaviest ones, it can mean the feature costs more than the plan.

The total LLM bill can't tell you which situation you're in. A bill that's 5% of revenue can hide a group of customers you lose money on every month.

This post is part of the "Where this leads" section of the complete guide to LLM cost optimization.

The three numbers

Cost per request. Input and output tokens at the model's rates, plus whatever your requests carry: system prompt, history, retrieved context, reasoning. How to calculate your real LLM API cost covers what goes in.

Requests per user per month. From your usage data, as a distribution. The median user and the 95th-percentile user often differ by ten times or more.

Revenue per user. The plan price, or what the customer actually pays after discounts.

Margin after LLM cost = (revenue − LLM cost) ÷ revenue. It's not gross margin yet; hosting, support and payment fees come off too. But it's the part that moves with usage, so it's the part to watch.

A worked example

A $29 plan with a chat feature at 1,200 input and 300 output tokens per request. These are illustrative usage numbers: a typical user makes 300 requests a month and a heavy user 3,000, and 10% of users are heavy. At each provider's published schedule on 17 Nov 2026:

ModelTypical userHeavy userHeavy user's margin
GPT-6 Luna$0.08$0.8197%
GPT-6.1 Terra$0.81$8.1072%
Claude Sonnet 5.5$1.62$16.2044%
Claude Opus 5.5$3.24$32.40loses $3.40

On Terra, the blended cost across all users is $1.54 a month, 5.3% of revenue. That looks comfortable, and it is, on average. The question is what happens to the heavy users when the model changes, the prompt grows, or a feature starts running an agent instead of a single call.

AI feature margin calculator

Enter your plan price and how a typical and a heavy user use the feature. The calculator shows LLM cost and margin for each, and blended across your users. Rates in USD per million tokens, checked 17 Nov 2026. LLM cost only: add your other costs on top.

Where AI unit economics go wrong

Heavy users on a flat plan. The most common case. A small group uses the feature many times more than the median. On a cheap model it doesn't matter; on an expensive one it decides the margin.

A model upgrade without a price change. Moving a feature to a bigger model can double or quadruple cost per request overnight, for every user, on the same plan price.

Agents and long conversations. Cost per request is no longer fixed. An agent run or a long chat grows with each step, so heavy users cost even more than their request count suggests. AI agent cost per run has the arithmetic.

Free users. They cost money and pay nothing. What a free user costs covers how to size and cap a free tier.

Price changes upstream. Gemini 3.6 to 3.8 Flash double in price on 1 January 2027. A margin worked out in December is wrong in January.

What to do with the numbers

  1. Find the customers you lose money on. Rank customers by LLM cost against what they pay. How to find unprofitable accounts walks through it.
  2. Cap or tier usage. Include a generous amount of usage in each plan and charge, or slow down, above it.
  3. Pick models per feature and per plan. The cheapest model that does the job on each feature, and possibly a different one for free or lower plans.
  4. Price for the heavy tail. Usage-based pricing, credits or a higher tier for heavy users all move the cost of the heavy tail onto the customers who create it.

Why AI SaaS gross margins run lower than classic SaaS covers the wider picture.

Measuring cost per customer

The provider bill gives one number for the whole account. To get unit economics you need the customer on every request.

LLMtrack records each request with its customer and feature, prices it at the model's rate, and shows cost per customer next to what that customer pays, so the accounts below your target margin stand out. You can start on the free plan.

import { fromOpenAIUsage } from 'llmtrack';

llmtrack.track({
  provider: 'openai',
  model: res.model,
  feature: 'chat-reply',
  customerId: account.id,
  customerName: account.name,
  ...fromOpenAIUsage(res.usage),
});

FAQ

Multiply your cost per request by the number of requests the user makes in a month. Use real usage data, and look at the median and heavy users separately rather than the average.

There's no universal number; it depends on your other costs and your pricing. Work out what margin you need after hosting, support and payment fees, and make sure your heavy users clear it, not just the average.

Usually, yes. Free users cost money and pay nothing, and a cap stops the heaviest ones without changing the experience for typical users.

Send a customer ID with every LLM request to your cost-tracking tool, and group cost by it. In LLMtrack that's the customerId field on each tracking call.

Because usage is uneven. The average hides a small group of heavy users whose LLM cost can exceed what they pay, especially on more expensive models or agent features.

Sources

Checked on 17 Nov 2026:

See margin per customer

Free plan, every feature, no card.

Start tracking free →