Clients usually ask what an AI feature costs to build. The better question is what it costs to run. The build is paid once; the AI bill shows up every month, and it scales with success. A feature that costs $40 a month at launch can cost $4,000 a month a year later if it wasn’t designed with costs in mind.
Here’s how AI pricing actually works, what typical features cost, and how to keep the bill sensible.
How AI Pricing Works in 60 Seconds
AI providers (OpenAI, Anthropic, Google, and others) charge per token — roughly ¾ of a word.
- Input tokens: everything you send — instructions, retrieved documents, chat history, the user’s message
- Output tokens: what the model writes back, usually priced several times higher than input
- Model tier: large frontier models cost far more per token than small, fast ones
The key insight: input usually dominates. A support bot that sends a 3,000-word policy document with every question is paying for 3,000 words every time — even if the answer is one sentence.
What Real Features Cost to Run
Typical monthly AI API costs for a small or mid-sized business, designed reasonably well:
| Feature | Typical volume | Monthly AI cost |
|---|---|---|
| Support chatbot (RAG over help docs) | 2,000–5,000 conversations | $30–$150 |
| Internal knowledge assistant | 20–50 staff, daily use | $50–$300 |
| Invoice / document extraction | 1,000–5,000 documents | $20–$150 |
| Lead qualification & email drafting | 1,000–3,000 leads | $20–$100 |
| AI voice agent | 2,000 minutes | $100–$400 (incl. voice + telephony) |
| Multi-step AI agent | 1,000 tasks, many calls each | $100–$1,000+ |
Hosting, database, and monitoring are separate — usually $50–$300/month for these projects. The full picture is in the web app cost breakdown.
Why Bills Blow Up
Too much context. Sending entire documents or full chat histories on every call.
One big model for everything. Using the most expensive model to classify an email or extract a date.
Agent loops. Agents that call the model 20 times per task, sometimes retrying the same failing step.
No limits. One user — or one bot — hammering the endpoint with no cap.
None of these are about the provider’s prices. They’re design decisions.
7 Ways to Cut the Bill
- Route by difficulty. Small, cheap models for classification, extraction, and routing; large models only for hard reasoning.
- Use prompt caching. Repeated system instructions and reference documents can be cached at a steep discount on most major APIs.
- Retrieve, don’t dump. Send the three relevant paragraphs, not the whole manual — this is what good RAG does.
- Trim chat history. Summarise older turns instead of resending them.
- Cap output length. Ask for concise, structured answers.
- Batch non-urgent work. Overnight document jobs can use discounted batch pricing.
- Set budgets and alerts. Per-user rate limits and a daily spend alert prevent surprises.
Applied together, these typically cut AI costs by 50–80% with no visible quality drop.
API or Self-Hosted?
For most businesses, APIs win: you pay only for what you use, and you get the best models without managing GPUs. Self-hosting open models makes sense for strict data-privacy requirements or very high steady volume — covered in the private AI comparison.
What to Ask Before You Build
- What’s the estimated AI cost per user or per task, not just per month?
- What happens to cost at 10× the volume?
- Which tasks use which model, and why?
- Is there caching, rate limiting, and a spend alert?
A developer who can’t answer these hasn’t designed for production yet.
Worried about what your AI feature will cost at scale — or already paying too much? Let’s talk. I’ll estimate running costs before you build, or audit an existing setup and show you where the savings are.