Why your bill is bigger than your estimate
Almost every AI chatbot budget is built from a single-turn estimate multiplied by conversation count. That undercounts by 2–5× because every turn resends the system prompt and the full prior conversation. A 10-turn chat with a 2,000-token system prompt sends 20,000 system tokens, not 2,000.
The three levers that actually move the number
In order of impact: cache the system prompt (75–90% off that portion), summarize history after 4–6 turns (routinely a 90% history-token reduction), and cap reply length (output is 3–5× the input price, so 300 tokens instead of 800 is the single biggest per-turn saving).
- • Cache the system prompt and any fixed retrieved context.
- • Summarize and discard raw history after 4–6 turns.
- • Cap max output tokens — and prompt for brevity, don't just truncate.
- • Route classification and routing steps to a cheap tier; reserve the premium model for the final answer.
Model against the 99th-percentile user, not the median
Median users are cheap; power users are what kill 'unlimited' plans. Run this calculator twice — once at your median conversation count and once at your 99th percentile — then price so the heavy tail is survivable. If the p99 user costs more than your plan price, you need credits, caps, or fair-use limits.
Related guides
Long-form playbooks on the same topic, written by the RevenueLab editorial team.
LLM Token Costs in 2026: Pricing Every Model, Hidden Multipliers, and Margin Math
Input vs output token pricing across GPT, Claude, and Gemini, the context-window cost trap, how caching and batching cut bills 40–80%, and the real per-user margin most AI apps miss.
Read the guideSaaS Pricing Strategy: Per-Seat, Usage, Tiers, and the Hybrid Future
A framework for choosing a SaaS pricing model — when per-seat caps your growth, when usage-based makes revenue volatile, and how hybrid models stitch the two together.
Read the guideFAQ
How much does it cost to run an AI chatbot?
For a typical support bot on a cheap/fast model — 6 turns, a 1,200-token system prompt, 300-token replies — expect roughly $0.005–$0.02 per conversation, or $125–$500/month at 25,000 conversations. Frontier models with long system prompts run 10–30× that.
Why does the system prompt cost so much?
Because it is resent on every single turn. A 2,000-token system prompt across 10 turns is 20,000 input tokens per conversation, often more than all user messages combined. Prompt caching is the fix.
Is caching worth setting up?
Almost always. Cached input typically bills at 10–25% of the normal input rate, and the system prompt is the most cacheable thing you have. On a prompt-heavy bot it routinely cuts total spend by 40–60%.
Should I use a cheap model or a frontier model?
Tier your pipeline. Route classification, intent detection, and retrieval decisions to the cheapest model; use the premium model only for the final user-facing generation. Production stacks that do this typically see 60–80% cost reduction with no perceived quality drop.
How do I convert words to tokens?
Roughly 1 token ≈ 0.75 English words, so multiply your word count by about 1.33. Code, non-Latin scripts, and heavy punctuation tokenize less efficiently — add 20–40% for those.
Do I need to include embedding costs?
Only if your bot retrieves documents. Embeddings are usually a rounding error at query time but can be significant at index time — model that separately in the RAG pipeline cost calculator.
How this calculator is built
Independently maintained
Written by Sam Doshi and the RevenueLab editorial team. We don't sell the data feeds this tool is built on.
Sourced from primary data
Benchmarks come from public AdSense / Stripe / IRS disclosures and reader-submitted data — never third-party "$X per view" claims. Full methodology.
Last editorial review
Reviewed on a rolling quarterly cycle. Dated reviews are published on the methodology record for each calculator.
Editorial standards
See our editorial policy and disclaimer. Results are estimates, not advice.