AI Chatbot Cost Simulator
Building a customer-support bot, an in-app assistant, or an internal AI tool? Estimate the monthly model bill from your real usage shape — and see how caching and model choice change it.
Input = system prompt + retrieved docs + history. A stable system prompt and knowledge base makes 40–70% of input cacheable at ~10% price.
Cost at your usage, across models
| Model | Monthly cost | Per user | Per 1,000 messages |
|---|
Standard-tier prices, August 2026 — see the AI price tracker. Batch processing (~50% off) applies to non-real-time workloads, not live chat.
The three levers that dominate chatbot cost
- Model tier. The spread between a budget and a frontier model is 25–100× per token. Most support bots route 80% of traffic to a cheap model and escalate the hard 20%.
- Prompt caching. Your system prompt and knowledge base are identical on every request — cached input bills at ~10%. For RAG bots this is routinely a 40–60% total saving.
- History management. Re-sending the whole conversation every turn is the lever people discover last and pay for first — worked out below, it turns a twenty-turn chat into roughly 210 messages' worth of input. Summarise or truncate beyond ~10 turns.
Sanity benchmarks
| Product shape | Typical model cost |
|---|---|
| FAQ/support bot, budget model, cached | $0.001–0.01 per conversation |
| General assistant, flagship model | $0.01–0.05 per conversation |
| Agent doing multi-step work | $0.10–1.00+ per task |
| Rule of thumb | If model cost per user exceeds ~10% of revenue per user, revisit routing and caching |
The cost that catches everyone: a chatbot re-sends the whole conversation
This is the single biggest reason a chatbot bill comes in above the estimate, and it is invisible in any per-message calculation, including the one above.
A language model has no memory between requests. For turn 12 of a conversation to make sense, the application has to send turns 1 through 11 along with it — every time. So the input tokens on each message are not the length of that message; they are the length of everything said so far.
Take a conversation where each turn adds about 300 tokens. Turn one sends 300. Turn ten sends around 3,000. Turn twenty sends around 6,000. Add them up and a twenty-turn conversation does not cost twenty messages' worth of input — it costs about ten times that, because the total grows with the square of the length rather than in step with it:
| Conversation length | Input tokens sent in total | vs. one message |
|---|---|---|
| 1 turn | ~300 | 1× |
| 5 turns | ~4,500 | 15× |
| 10 turns | ~16,500 | 55× |
| 20 turns | ~63,000 | 210× |
The estimate above prices messages, so it is accurate for a support bot that answers one question and stops, and low for anything where people hold a conversation. If your product keeps a thread going, multiply the input side by the average conversation length before you trust the monthly figure.
This is also why the two standard fixes exist. Prompt caching works well here precisely because the repeated part is enormous and identical. Trimming history — keeping the last few turns plus a short summary instead of everything — attacks the same growth from the other side, and usually costs less quality than people fear.
Frequently asked questions
What usage shape should I assume before launch?
A common planning baseline for consumer products: 10–20% of signups become monthly active, active users send 5–20 messages/month, and support-style conversations average 3–6 turns. Model with your best guess, then re-simulate with real data after two weeks.
Does this include infrastructure costs?
No — this is the model (LLM API) bill only. Hosting a thin chat backend typically adds $5–50/month at small scale; vector databases for RAG add more.
How do rate limits affect scale?
Entry API tiers allow thousands of requests/minute — enough for most products. Past that, providers raise limits with usage history or a sales conversation, not extra fees.
Should I use batch pricing?
Not for live chat (batch takes up to 24h). Use it for the offline parts: nightly summarization, embedding generation, analytics — those get ~50% off.
You might also need
🤖AI API Cost Calculator
Compare GPT, Claude & Gemini pricing and estimate monthly costs.
💠AI Pricing by Provider
Every model priced, compared within each provider lineup.
🔢Token Calculator
Token count and cost of any text, on every model.
⚖️Subscription vs API
Find your break-even between $20/mo plans and pay-per-use.