How Much Does an AI Chatbot Cost for 10,000 Users?
A worked monthly bill for 200,000 chat replies on every tracked model, the cost per user, and four ways to cut it.
Updated September 19, 2026
“What will this cost at scale?” is the question every AI feature runs into before launch. This guide prices a concrete case: a chat assistant used by 10,000 people a month, each sending about 20 messages. That is 200,000 replies a month. Every number below comes from the providers’ official list prices.
The assumptions
- 1,200 input tokens per message. A system prompt of roughly 600 tokens plus the recent conversation history.
- 350 output tokens per reply, around 250 words.
- Half the input served from the prompt cache. The system prompt repeats on every request, so it is cached; the conversation is not.
Change any of these and the totals move proportionally — that is what the cost calculator is for.
Monthly cost for 200,000 replies
| Model | Provider | Per reply | Per user | Per month |
|---|---|---|---|---|
| GPT-5.6 Luna | OpenAI | $0.0006 | $0.01 | $110.40 |
| DeepSeek V4.1 Flash | DeepSeek | $0.0006 | $0.01 | $120.72 |
| Gemini 3.5 Flash-Lite | $0.0011 | $0.02 | $214.60 | |
| Gemini 2.5 Flash | $0.0011 | $0.02 | $214.60 | |
| Grok 4.3 | xAI | $0.0017 | $0.03 | $349.00 |
| Gemini 3.8 Flash | $0.0018 | $0.04 | $361.50 | |
| DeepSeek V4 Pro | DeepSeek | $0.0022 | $0.04 | $440.88 |
| Claude Haiku 4.5 | Anthropic | $0.0024 | $0.05 | $482.00 |
| Grok 4.6 | xAI | $0.0036 | $0.07 | $720.00 |
| Gemini 3.5 Flash | $0.0041 | $0.08 | $828.00 |
The spread is the story. GPT-5.6 Luna runs the same traffic for $110.40 a month, while a flagship model like Claude Sonnet 5 costs $964.00 — a 9x difference for identical volume.
Four ways to cut the bill
1. Cache the system prompt
Cached input tokens cost a fraction of fresh ones. In this workload, caching half the input saves 18% on Claude Sonnet 5 — $1,180 without caching versus $964.00 with it. The catch is that caches expire after minutes of inactivity, so the saving is real only when traffic is steady.
2. Shorten the system prompt
Every token in the system prompt is billed on every single message. Trimming 300 tokens from it removes 60 million tokens a month from this workload.
3. Cap the reply length
Output tokens cost several times more than input tokens on every provider. Asking for concise answers, or setting a max output limit, is usually the fastest saving available.
4. Route by difficulty
Most support questions are easy. Sending them to a budget model and escalating only the hard ones to a flagship gives you most of the quality at a fraction of the price. Compare capability against price on the benchmarks page before you pick the two ends of that route.
What this guide does not include
Only model tokens. A production chatbot also pays for embeddings or search, hosting, logging and moderation, plus any failed requests — a retry is billed like any other call. Budget a margin on top of the numbers above.
Next: compare two models side by side or check the full price table.