How does Claude API pricing work for a small team?

Claude API pricing is a meter, not a seat fee. Anthropic charges for tokens sent to and generated by each model, with additional lines for writing and reading cached material. Estimate a specific workflow from measured token counts, then apply the model’s current rates. A Claude subscription is not API credit.

Four lines make up the Claude token bill

An API, or application programming interface, lets a team’s software send a request to Claude. The Anthropic price page quotes each rate per one million tokens. A token is a small unit of text, not exactly a word; our token guide explains the count.

Claude’s table has four lines worth separating. Input tokens are new instructions, conversation and source material. A cache write stores eligible input for reuse. A cache hit, also called a cache read, retrieves that stored material at a discount. Output tokens are the response Claude generates. A repeated 50-page handbook may therefore create a higher write charge once and a much lower read charge on later questions, provided the requests meet Anthropic’s caching rules.

On Anthropic’s standard table fetched on 7 September 2026, Claude Haiku 4.5 costs $1 per million fresh input tokens, $1.25 for a five-minute cache write, $0.10 for a cache hit and $5 per million output tokens. Claude Sonnet 5 costs $2, $2.50, $0.20 and $10 on the same four lines. Claude Opus 5 is $5, $6.25, $0.50 and $25. The chart shows fresh-input rates for all Claude models in our dated API pricing ledger. It does not fold cache or output prices into a misleading single number.

Anthropic also lists a one-hour cache-write option at a higher rate than a five-minute write. The right duration depends on how soon the same prefix returns. A prefix is the unchanged beginning of a request. If a team asks many questions about the same handbook throughout an hour, paying to retain that prefix can be useful. If the material is never reused, the write is an extra charge with no later saving.

Work out a monthly cost from one workflow

Begin with observed use. Suppose an internal document assistant handles 8,000 requests in a month. Each request has 5,000 input tokens and produces 500 output tokens. That is 40 million input tokens and 4 million output tokens. Without caching, the model-only formula is:

40 x fresh input rate + 4 x output rate

At the published Claude Haiku 4.5 rates, the result is $40 plus $20, or $60. At Claude Sonnet 5 rates, it is $80 plus $40, or $120. These are worked examples from the standard rate card, not invoices. They exclude retries, taxes, special features and any cache writes.

For a cached estimate, divide input into changing and reusable portions. If most of the request is a stable handbook, count the first eligible request as a cache write and later eligible requests as cache reads. Keep the changing instructions on the fresh-input line. Do not simply mark the same share of all future input as cached. Anthropic’s response reports actual cache usage, and that measured field should replace the estimate after a pilot.

Use Anthropic’s token-counting endpoint before large calls. It estimates input that includes system instructions, messages, tools, images and PDF documents. A system instruction is a standing rule sent before the user’s message. A tool definition describes another function Claude may call. Both consume input even when a user never sees them in the chat. Anthropic notes that a preflight count is an estimate, so leave room near a context or budget boundary.

Limits govern speed and total spend

Anthropic documents several API limits. Requests per minute restrict how many calls begin. Input tokens per minute restrict incoming volume, and output tokens per minute restrict generated volume. The service uses a token-bucket method, meaning capacity refills over time rather than resetting only at a clock boundary. When traffic exceeds a limit, software receives a 429 response with retry guidance.

A spend limit is different. It caps accumulated API cost for an account tier. Traffic can be slow enough to satisfy every rate limit and still reach the spend limit by continuing all month. Monitor both. Queueing and retry logic handle short bursts; budget alerts and spend controls handle the financial ceiling.

Consumer Claude usage is a third system. The Claude plans page describes chat subscriptions, while the plan comparison records their published allowances. Those rolling chat limits do not define API throughput, and Pro or Max subscription fees do not pay an API invoice. Keeping the systems separate prevents a common small-team budgeting error.

Choose a Claude model with evidence

Price is useful only after a model passes the task. Build a test set, a fixed group of real examples with an agreed scoring rule. Test Haiku, Sonnet and Opus on the same prompts. Record result quality, input tokens, output tokens, cache use and response time. Select the lowest-cost model that reliably clears the required score.

This matters because a cheap call that needs repeated correction can cost more than a successful dearer call. The reverse is also true: using the most expensive model for simple extraction pays for ability the workflow does not use. The Anthropic hub connects the company’s model, plan and status information, and our guide to reading model announcements shows how to separate a benchmark claim from evidence about your own task.

What it means for you

Run a small pilot before approving a monthly number. Log the exact Claude model, fresh input, cache write, cache read and output counts for each workflow. Price typical use and a busy month using the current primary page, then add room for retries. Give development and production separate keys or workspaces so an experiment cannot consume the live budget unnoticed.

Review cache performance rather than assuming it. Review limits as both speed and spend controls. Most of all, put a date beside every rate. Anthropic can change prices, models and limits, so the sound estimate is one that another person can reproduce from the source and the recorded usage.