How to Estimate Your Monthly LLM Token Usage
Every LLM budget starts with one number: tokens per month. This guide turns things you actually know — requests per day, words per prompt, words per reply — into that number, and covers the hidden tokens that make real bills two or three times the naive estimate.
From words to tokens
A token is a chunk of text roughly three-quarters of an English word, or about four characters. You do not need a tokenizer to budget: multiply words by 1.3 and you are within ~10% for English prose. From there, the estimate is three measurements and one multiplication:
- Measure the average input per request: the user's message plus everything you send with it (system prompt, retrieved context, chat history). Count words, multiply by 1.3.
- Measure the average output per request: the model's reply, same conversion.
- Count requests per day — from analytics, or requests per user × active users.
- Then:
tokens/month = requests/day × 30 × (input tokens + output tokens).
The hidden tokens that break estimates
The user-visible text is usually the smallest part of a request. Four things quietly multiply your volume. A system prompt is re-sent with every single request — a 1,500-token prompt at 36,000 requests a month is 54M tokens before a user types anything. Chat history grows with every turn and is re-sent in full each time, so a 20-turn conversation costs far more than 20 single questions. RAG context (retrieved documents) commonly adds 1,000–4,000 input tokens per request. And tool definitions for function calling ride along on every call too. This is also why the input:output ratio in cost calculators matters: production chat and RAG workloads typically send 3 to 6 input tokens for every output token.
Worked example: a support chatbot
Worked example. A support bot handles 1,200 messages a day. Each request carries a 1,500-token system prompt + retrieved context, plus a 100-token user message (1,600 input tokens), and produces a 250-token reply. Monthly volume: 1,200 × 30 × (1,600 + 250) = 66.6M tokens/month at a 6.4:1 input:output ratio. On GPT-5.4 Mini ($0.75 in / $4.50 out per 1M) that is 57.6M × $0.75 + 9M × $4.50 ≈ $84/month — of which almost half is the re-sent system prompt.
Rules of thumb
| Content | ≈ Tokens | Notes |
|---|---|---|
| 1 English word | 1.3 | ≈ 4 characters per token |
| One A4 page (500 words) | 650 | prose; tables and code run denser |
| 10-page PDF in a RAG context | 6,500 | per request it is retrieved into |
| Typical system prompt | 200–2,000 | re-sent with every request |
| Typical chat reply | 150–400 | output — the expensive direction |
| 1,000 lines of code | 10,000–15,000 | code tokenizes denser than prose |
| Non-English text | +20–100% | varies by language and script |
Two more corrections before you trust the number. Anthropic's newest models (Fable 5, Opus 4.8, Sonnet 5) tokenize the same text into roughly 30% more tokens, so multiply volume accordingly when comparing providers. And leave 30–50% headroom for growth and retries — bills only ever surprise in one direction.
Got your tokens-per-month number?
Drop it into the calculator to compare API prices and see where self-hosting breaks even.
Open the LLM Token Cost Calculator