How Much Does an AI Chatbot Cost to Run?
Search this question and you'll find two kinds of answers: SaaS pricing pages ($50–500/month for a widget) and per-message math that ignores how LLM billing actually works. This guide does neither — it prices a realistic support conversation on live rates (updated 2026-08-23), shows the resend mechanic that quietly inflates real bills past naive estimates, and gives monthly numbers for three honest sizing tiers. Want your own workload instead? The chatbot cost calculator runs this exact math interactively, including caching effects.
The unit that matters: cost per conversation
Our reference conversation: 8 turns, an 800-token system prompt, 60-token user messages, 180-token replies. Because the API is stateless, each turn re-sends the system prompt plus everything said so far — so the average turn processes ~1,700 input tokens, not 860:
| Model tier | Per conversation (real) | Per conversation (naive per-message math) | Underestimate |
|---|---|---|---|
| OpenAI GPT-5.4 Mini (mini) | $0.0167 | $0.0116 | ×1.4 |
| OpenAI GPT-5.6 Sol (flagship) | $0.0416 | $0.0282 | ×1.5 |
That ×1.5 gap is the single most common budgeting mistake — the mechanics are unpacked in how LLM API pricing works.
Monthly bills at three honest sizes
| Tier | Conversations/day | On GPT-5.4 Mini | On GPT-5.6 Sol |
|---|---|---|---|
| Side project | 50 | $25.02/mo | $62.40/mo |
| Growing product | 500 | $250/mo | $624/mo |
| Busy support desk | 5,000 | $2502/mo | $6240/mo |
Read the columns, not the rows: at every size, the model tier decides the bill. A busy support desk on a mini model costs less than a growing product on a flagship. Most routine support conversations do not need flagship reasoning — and the cheapest-APIs ranking shows how far down the price floor goes if your workload is simple.
What's not in these numbers
- Caching, working for you: the static system prompt is a perfect cache candidate — 75–90% off its repeated cost on supported models (see the caching guide). Our calculator models per-provider cache TTLs, including the hit-rate cliff when users reply slowly.
- Retrieval: if the bot answers from your docs, add RAG token costs — the RAG calculator prices that pattern.
- The wrapper: hosting, the chat widget, logging, and human-handoff tooling — near-zero on a static-page bot, $50–500/month if you buy a platform for them.
- Voice: a spoken bot adds STT + TTS per minute, which usually exceeds the LLM line — the voice agent pricing guide covers that stack.
Budget in four steps
- Estimate conversations/day honestly (support tickets ÷ deflection rate is a good proxy).
- Price it on a mini tier first in the calculator — escalate to flagship only the conversations that fail.
- Multiply by 1.2 for retries, abandoned sessions, and prompt experiments.
- Re-check monthly: prices only move down, and the tier gap keeps widening — this page's numbers refresh daily.