Cost by feature type, with the math shown
All figures at current rates (live at useweckr.com/pricing.json), on gpt-5.4-mini at $0.75 input and $4.50 output per million tokens unless noted:
- Single shot transforms (summarize, classify, extract, rewrite): 2,000 tokens in, 350 out is about $0.003 per call. At 160 calls per active user per month: about $0.49 per user. The full worked version is in estimating AI feature cost.
- Generation with a cap (drafts, emails, descriptions): output dominates, so
max_tokensis the price dial. A 1,000 token draft costs about $0.005 on mini; typical usage lands under a couple of dollars per user. - Chat: the deceptive one. Each turn resends the conversation, so a 20 turn conversation costs several times its final answer alone. Realistic monthly range on a mid tier model: $1 to $15 per active user, tail higher. The compounding math is in AI chatbot cost per conversation.
- Agents and multi step workflows: context compounds across steps and users vary 100x. Budget by measurement, not estimate, and read AI agent cost control before shipping one.
Why the estimate is not the plan
Every number above is a typical case, and AI usage is heavy tailed: a small fraction of users drive a large share of cost, and one power user can cost more than their plan price, invisibly inside a healthy average. The estimate answers “can we afford to ship it”. The plan needs two more pieces: per user measurement from week one, and per user spending caps so the tail has a ceiling. Your plan pricing then gets checked against the measured distribution, the audit in is your unlimited AI plan profitable.
The four affordability levers
- Cap output. Output tokens cost about 5x input; a tight
max_tokenstrims every call invisibly. - Route down. Budget models handle most product features; save the premium models for the tasks that fail without them, per model routing for LLM apps.
- Cache the stable context. Repeated system prompts at a tenth of the fresh rate, with the write premium caveat in prompt caching economics.
- Cap the users. Downgrade or block past a monthly budget, so no single account can turn your feature into a loss.
Try your own numbers in the break even calculator: plan price, calls per user, model, and it shows margin per user and your break even point.
FAQ
How much does it cost to add AI to an app?
Development cost aside, the API spend for a typical single feature is far lower than founders expect: a summarizer or writing assistant on a budget model runs roughly $0.25 to $1.50 per active user per month at normal usage. A chat feature runs higher, roughly $1 to $15 per active user depending on conversation length, because context resends every turn. Agentic features are the expensive tier, and the heavy tail of users dominates all of these averages.
What does an AI chatbot cost per user per month?
On a mid tier model like claude-sonnet-4-6, a user having a few short conversations a week costs roughly $1 to $3 a month, while a daily heavy user with long conversations can reach $10 to $30, because each turn resends the accumulated conversation. The per conversation math and the context compounding are the whole story, and they are why chat features need per user caps more than any other feature type.
Which AI feature types are cheapest and most expensive?
Cheapest: single shot transforms with bounded output, classification, extraction, short summaries, often fractions of a cent per call on budget models. Middle: generation features with capped output. Expensive: chat (context grows per turn) and agentic workflows (multi step loops resend context each step and can vary 100x between users). Price the feature type before choosing it, not after.
How do costs change as my app grows from 100 to 10,000 users?
Roughly linearly in active users, with two corrections: the heavy tail widens (your p95 user gets more extreme as the population grows), and per user variance means the average misleads. At 100 users a bad estimate stings; at 10,000 it is a five figure surprise. The fix is the same at both scales: measure cost per user from the first week, and cap the tail.
How do I keep an AI feature affordable after launch?
Four levers, in order of ease: cap output tokens (output costs 5x input), route to the cheapest capable model, cache stable prompt context, and set per user spending caps that downgrade or block runaways. Then watch cost per user against plan price continuously, because usage patterns and provider prices both drift. That watching layer is what Weckr provides in two lines.
Keep reading
Ship it cheap. Keep it cheap on purpose.
The feature will cost less than you feared at launch and more than you noticed by month three, unless cost per user is visible the whole way. Weckr makes it visible in two lines, with the caps that keep the tail bounded, free for 50,000 requests a month. Click through the live demo, or start with the AI cost and margin guide.