How to calculate LLM cost per user (formula, examples, benchmarks)

The formula nobody writes down

Every credible LLM cost-per-user number reduces to the same equation. Memorise it once and you can sanity-check any AI feature on the back of a napkin:

Four inputs, one output. The trick is that three of those four inputs are wrong in most spreadsheets we see — and the answer is extremely sensitive to all of them.

Step 1: Get the per-call cost right

Per-call cost is where almost everyone underestimates. Output tokens cost 3-5× more than input tokens on most frontier models, so a chatty, long-answer feature will blow past a "summarise" feature even when input lengths are similar. Below is the live shape of the market in May 2026, in dollars per 1M tokens.

Step 2: Multiply by realistic usage, not best-case

Two numbers matter here: calls per active user per day and active days per month. The founder mistake is plugging in your power user's behaviour as the average. The fix is borrowing the analytics convention: your blended usage is closer to the median than the mean.

• AI chat assistants (Intercom Fin, Notion AI): 3–8 calls/day per active user, 12–18 active days/month.
• Workflow automations (Zapier AI, Bardeen): 15–40 calls/day, 8–14 active days/month.
• Coding copilots (Cursor, Copilot): 50–200 calls/day, 18–22 active days/month.
• Agentic apps (Devin-style, browser agents): 5–20 calls/run × 1–3 runs/day, but 5,000+ tokens per call.

Step 3: A worked example you can copy

Suppose you run an AI writing assistant. Average prompt is 500 input tokens, response is 800 output tokens, the average user fires 6 calls on each of 15 active days/month. On GPT-4o:

Step 4: See the curve, not the point

A single per-user number is useful for unit economics, but founders also need the total bill at scale. The chart below shows monthly AI spend across MAU, holding the writing-assistant usage profile fixed and varying only the model.

The four numbers most founders get wrong

1. Output tokens. Engineers estimate based on "what feels short". Pull a week of real responses from your logs and average the actual completion lengths.
2. System prompts and tool context. Your "500-token user prompt" is actually 500 + 1,800 tokens of system instructions, function schemas, and retrieved documents on every call.
3. Retries and tool loops. Agent-style features that retry on validation errors or run multi-step tool chains routinely double the call count.
4. Power-user concentration. The top 5% of users consume 40-60% of calls. A blended per-user number hides whether your top decile is unprofitable.

Where to go next

Once you know per-user cost, the next question is how to size your monthly AI spend across features and forecast it through a fundraising milestone. We cover that in our companion guide: AI API budget planning for startups. If you only have ten minutes today, run your numbers through the SaaS margin simulator and see whether your AI feature actually clears 60% gross margin.