Count Tokens and Estimate Your LLM API Costs Before You Ship

# What this guide explains

# What a token actually is A token is a chunk of text the model treats as one unit. Common words are often one token. Rare words, punctuation, spaces, line breaks, JSON, code, UUIDs and other identifiers can split into many tokens. Tokenization depends on the content, not simply length. That is why rules of thumb based on characters or words drift widely across real inputs.

# The rule of four characters per token is unreliable

# How to count reliably

# Watch out for envelope tokens Counting only visible message text undercounts. Real API requests include structural tokens for roles and message framing. Measured requests show roughly seven to nine tokens of envelope per request. That overhead is small for very large prompts but dominates the cost of many tiny requests.

# Output costs and cost shape Providers charge more for generated output tokens than for input tokens. Across major providers measured, output costs were two to six times higher than input. That means a system that produces long replies will have a different cost profile than one that performs short summaries or tagging. Measure both input and expected output token counts to model costs accurately.

# Conversation history and the common surprise Teams commonly discover large bills when a system keeps resending conversation history. Ten conversation turns can represent about 5.1× the text that actually exists in a single turn, because history accumulates every request. This growth is the most frequent source of unexpected spend.

# Prompt caching and discounts Prompt caching is an available discount: cached reads are billed at roughly a tenth of normal input costs, according to measured behavior. Prompt caching is the largest effective discount available, but many systems do not use or claim it. Where caching makes sense, it sharply reduces per-request input cost.

# Practical next steps 1) Tokenize representative prompts, system messages, and likely user inputs with the same tokenizer the model uses. 2) Measure input and output token counts separately. 3) Include envelope tokens per request in your model. 4) Model conversation growth over expected turns to estimate accumulation. 5) Implement prompt caching where appropriate to cut input cost.

# Final point The usage field returned by API responses provides exact billed counts. After one real call, you can replace estimates with measured values and budget accurately.