LLM Monthly Cost Calculator
Turn DAU, requests per user and token sizes into a monthly LLM bill. Compare models on one chart, with prompt caching and multi-turn chat. In-browser.
List prices as of 2026-09-02. Caching applies to the prompt prefix only — the way real prompt caches work.
Service type
Replaces the token profile only — your traffic, models and budget stay.
Traffic
Request profile
Models to compare
Anthropic
OpenAI
Result
for GPT-5.4List prices only — free tiers, promotional credits, committed-use and enterprise discounts are not included.
Retries, image and audio input, tool execution, embeddings and vector databases are not modelled. Leave about 20% headroom.
This tool calculates cost only. For speed (tokens/sec, TTFT) and quality see Artificial Analysis. artificialanalysis.ai
How to use
- Pick a service type — chatbot, summarisation, RAG, coding agent…
- Set your DAU and how many requests each user makes per day.
- Choose up to eight models to compare on the chart.
- Read the monthly bill, cost per user, and share the link as your estimate.
FAQ
How much does a chatbot with 1,000 daily users cost?
With the built-in chatbot profile — 500-token system prompt, 150 tokens per user turn, 400 per reply, six turns per session, two sessions per user per day — 1,000 DAU is 60,000 sessions a month. That comes to about $355 on GPT-5 mini with prompt caching on ($0.36 per user), $518 on Gemini 2.5 Flash and $2,831 on GPT-5.4. Turning caching off pushes the GPT-5 mini figure to $470.
How is cost per user calculated?
Monthly bill ÷ DAU. The monthly bill itself is DAU × requests per user per day × days in the month × cost of one request, so the per-user figure is simply requests per user per day × days × cost per request — it does not depend on how many users you have. That is why it is the number to quote in a business case: it stays true as you grow.
Why does a longer conversation raise the cost per session so steeply?
Every turn re-sends the whole history, so the input tokens billed in one session grow with the square of the turn count: N·S + N·U + (U+A)·N(N−1)/2. Doubling six turns to twelve does not double the session price — it roughly triples it. The “Conversation turns” chart draws that curve, and the dashed line shows what prompt caching takes back.
How much does prompt caching actually save here?
This calculator applies the cache hit rate to the prefix only — the system prompt plus any fixed context — because that is the part a prompt cache can match. The user’s own message is new every time and is always billed at the full input price. Most calculators multiply the whole input by the hit rate, which under-states the bill. On the chatbot profile the honest saving is about 25%; on the coding-agent profile, where the prefix is 22,000 tokens, it is over 50%.
Where does the Korean won figure come from?
The won line uses the European Central Bank reference rate, fetched read-only in your browser and cached locally — not a fixed rate baked into the page. If the rate cannot be fetched the won line simply disappears and the dollar figures stay. Nothing you type is ever sent anywhere.
What is not included in this estimate?
List prices only: free tiers, promotional credits, committed-use and enterprise discounts are not modelled, and neither are retries, image or audio inputs, the cost of running the tools a model calls, embeddings, vector databases or your own servers. Speed and quality are out of scope too — check Artificial Analysis for those. Treat the result as a floor and add about 20% headroom.
Related tools
More AI & Data tools: AI & Data