Cost Calculators → Cost Scenarios

RAG Chatbot — 10,000 Monthly Active Users

Monthly LLM API cost for a RAG-powered chatbot serving 10,000 MAU, with typical conversation patterns.

Low estimate
$5,220/mo
Mid estimate
$7,400/mo
High estimate
$11,000/mo

Assumptions

  • 10,000 MAU, avg 8 conversations/user/month = 80,000 conversations
  • Avg 6 turns per conversation
  • System prompt: 1,500 tokens
  • RAG context: 3 chunks × 500 tokens = 1,500 tokens
  • User message: 150 tokens avg
  • Assistant response: 400 tokens avg
  • Total input per turn: ~3,150 tokens (system + RAG + history + user)

Base model: Claude Sonnet 5 — $3/M input · $15/M output · 2026-08-19

Cost breakdown by component

System prompt (×all turns)12%
RAG chunks (×all turns)28%
Conversation history38%
User messages8%
Assistant outputs14%

How to cut this cost by 50%+

Enable prompt caching
Cache system prompt + static RAG context with Anthropic 5m TTL
25-35% savings
Sliding window (5 turns)
History cost is quadratic — capping at 5 turns drastically reduces it
30-40% savings
Route simple queries to Haiku
FAQ and simple Q&A can use Claude Haiku 4.5 at $0.80/M
40-60% savings
Calculate for your exact workload

These are estimates based on typical patterns. Enter your specific token counts and volume for a precise calculation.

Open calculator →conversation cost calculatorprompt caching calculator

Frequently Asked Questions

What does a RAG chatbot cost for 10,000 users?

Roughly $5,000–$11,000/month on Claude Sonnet 5, depending on conversation length and RAG chunk count. The biggest lever is conversation history — each additional turn adds all previous turns to input. Enable prompt caching and limit history to 5-7 turns to reduce costs by 50%+.

What is the biggest cost driver in a RAG chatbot?

Conversation history, which grows quadratically per conversation. By turn 6, you're resending 5 full turns of dialogue as context. The second driver is the number of RAG chunks per query — retrieving 10 chunks instead of 3 triples that cost line.

Pricing verified 2026-08-20. All figures are estimates with stated assumptions. See methodology.