Every price on every Claude platform — Bedrock, Vertex AI, Foundry, Claude Platform on AWS, or direct from Anthropic — is quoted per million tokens. So before any budgeting, capacity planning, or vendor conversation makes sense, you need an intuition for what a token is and roughly how many your workload consumes. The good news: the intuition takes five minutes to build and never changes.
What a token actually is
Language models don't read text letter by letter or word by word. Text is split into tokens — chunks that are usually whole short words or pieces of longer ones. "The" is typically one token; "internationalization" splits into several. Punctuation, spaces, numbers, and code all tokenize too, and non-English languages often use more tokens for the same meaning than English does.
You never handle tokens directly. You send ordinary text; the platform tokenizes it, the model responds, and the response comes back as text along with a count of tokens consumed in each direction. Those counts are the entire basis of your bill.
Input and output are priced differently
Every request has two meters. Input tokens are everything you send: the user's question, your system instructions, and any documents or conversation history you include. Output tokens are everything Claude writes back. Output is consistently the more expensive direction — on Claude Opus 4.8, list price is $5.00 per million input tokens versus $25.00 per million output tokens; on Claude Haiku 4.5 it is $1.00 versus $5.00.
This asymmetry shapes real bills in a way that surprises newcomers twice, in opposite directions. Summarization workloads (huge input, short output) are cheaper than the document sizes suggest. Generation workloads (short brief in, long report out) are more expensive than the tiny prompt suggests. Neither is a problem; both are worth knowing before you promise a budget number.
Working a real example
Take a customer-support assistant. A typical exchange might include a page of system instructions and retrieved policy text (~700 tokens), the customer's message (~100 tokens), and a two-paragraph reply (~200 output tokens). That's roughly 800 input and 200 output tokens per exchange. At Claude Sonnet 5 list prices ($3.00 input / $15.00 output per million tokens), one exchange costs a fraction of a cent, and even 100,000 exchanges a month stay in the hundreds of dollars — before optimizations like prompt caching. The arithmetic is deliberately simple: volume × tokens per interaction × price per token. Anyone in finance can audit it.
Two multipliers deserve attention early. Conversation history: chat applications typically resend prior turns as input with every new message, so long conversations get progressively more expensive per turn. Repeated boilerplate: if every request carries the same long instructions or reference documents, prompt caching — available on all four platforms — lets you avoid paying full price for the repeated portion, and is usually the first optimization worth adopting.
Tokens also bound what fits in a request
Beyond billing, tokens measure capacity. Each model has a context window — the maximum tokens a single request can span, prompt and response together. Claude Opus 4.8 and Sonnet 5 take up to 1M tokens; Claude Haiku 4.5 takes 200K. A million tokens is enormous — on the rough page math above, over a thousand pages — but "it fits" and "it's free" are different statements: a maximally stuffed request is also a maximally priced one. Send what the task needs, not everything you have.
Where to go next
Turn intuition into a budget with estimating your monthly Claude bill, and see the biggest single lever in prompt caching. Current per-model prices live in the model lineup explained.