How Much Does an AI App Cost to Run Per User?

The cost of running an AI app comes down to one formula: AI cost per user per month = AI actions per user × tokens per action × price per token. A user who runs 200 actions a month, at about 2,000 tokens each, on a model that costs 3 per million tokens, costs you about 1.20 a month.

That number sounds small, and for many apps it is. The trouble starts when it's multiplied by heavy users, long conversations and a free plan. Let's build it properly.

What a token is

AI models charge by the token, a small chunk of text. A common rule of thumb for English is that one token is roughly four characters, or about three-quarters of a word. So 1,000 tokens is about 750 words.

Two details matter for cost:

• You pay for input and output. The text you send (your instructions, the user's message, any documents or chat history) and the text the model writes back are both counted.
• Output usually costs more than input. Most providers price them separately, so a feature that writes long answers costs more than one that reads long documents and replies briefly.

Prices differ a lot between models and change often, so always use your provider's current pricing page.

The formula, step by step

1. Actions per user per month. How many times does a typical user trigger the AI? Count generations, chat messages, summaries, whatever calls the model.
2. Tokens per action. Add the input (system prompt + user text + context) and the output. If input and output have different prices, work them out separately.
3. Price per million tokens. Use a blended price if you mix input and output: for example, if three-quarters of your tokens are input.
4. Multiply. Actions × tokens × price ÷ 1,000,000.

Here's the example with separate input and output prices (made-up prices for illustration):

At 200 actions a month, that's 200 × 0.006 = 1.20 per user.

The hidden multipliers

The happy-path number is almost always too low. Watch for these:

• Chat history. Every new message in a conversation often resends earlier messages as context. The tenth message can cost several times more than the first.
• Long system prompts. Instructions you send with every call add up across thousands of calls.
• Retries and failures. Timeouts, reruns and "regenerate" buttons all bill.
• Documents and retrieval. Pasting a long document into the prompt can dwarf the user's question.
• Heavy users. Usage is rarely even. A small group of power users can generate a large share of your costs.

A safer planning habit: estimate the average, then double it for the first version of your plan. Tighten it once you have real usage data.

Light, typical and heavy users

Averages hide the users who cost you most. Before you set prices, split your users into three groups and cost each one:

The heavy user costs 100 times more than the light one. If one customer in twenty is heavy and the rest are typical, that one customer costs almost as much as the other nineteen together. That's exactly what monthly limits, credit packs and bring-your-own-key options are for.

A practical way to estimate tokens for a new feature: run it twenty times with realistic inputs, log the input and output token counts your provider reports, and use the average plus a margin. It takes an afternoon and replaces guesswork with data.

How AI cost should shape your pricing

AI cost is a variable cost, so it lowers your gross margin with every active user. A few rules keep it under control:

• Keep it a small share of revenue per user. If a customer pays 12 a month and costs you 1.20 in AI, that's 10%. If they cost 4.00, a third of your price is gone before hosting, fees and support.
• Put limits on free plans. A generous free plan with unlimited AI is an open tab.
• Be careful with lifetime deals. A lifetime buyer pays once and uses AI every month. Read how to price a lifetime deal before you offer one.
• Consider credits. Monthly credit allowances with paid top-ups tie your cost to your revenue.

Ways to cut AI cost per user

• Use smaller models for simple tasks. Classification, formatting and short summaries rarely need your most capable model.
• Trim prompts. Shorter instructions and less repeated context cut input tokens on every call.
• Summarise long conversations instead of resending the full history.
• Cache repeated work where the same input produces the same answer.
• Let users bring their own key (BYOK). Some users, often power users, prefer to pay the provider directly. Their AI cost to you drops to zero.
• Set sensible limits per plan, and show users how much they've used.

Track it every month

Your plan's AI cost is an assumption until you have real bills. Once you launch, compare the provider's invoice with your active users each month. If the real cost per user is higher than planned, update the assumption before it surprises you in your cash flow.

Work out yours

Our free AI cost per user calculator shows what each user costs you in AI every month, and how many months one lifetime sale pays for. In a full startzero.money plan, AI cost flows straight into your gross margin, break-even point and runway, so a price change at your provider shows up everywhere at once.