Anthropic API
by Anthropic PBC
Getting-started guide
The same models as the chat assistant, billed per token — and under a contract with the opposite privacy default.
Current prices: $0.10 input / $0.50 output per million tokens · See all plans
Last reviewed 2026-10-14 · for Anthropic API
Your own code sends text to Anthropic's models and gets a response back, instead of you typing into a chat window. You create an account on the developer console, generate a key, and your program sends requests with that key attached. You read what comes back, change the request, and send it again.
A pay-per-use interface to the same model family behind the Claude chat assistant.
Aimed at people building something — a tool, a script, a product feature — rather than at people who want to chat.
Billed by usage with no subscription floor, which means near-zero cost while you experiment and real cost at volume.
Governed by commercial terms, which have a different and stricter privacy default than the consumer chat product.
You need to write or run code. There is no interface here — the console has a testing workbench, but the product is the endpoint.
Pricing is per million tokens and differs sharply by model: the largest models cost many times what the smallest do, and the text a model writes back costs more than the text you send it. Anthropic's pricing page lists the current rate for each model, including older ones.
Batch processing halves the cost in both directions for work that does not need an immediate answer. If your job can wait, this is the single largest saving available.
Set a spending limit in the console before your first real run. Runaway loops are the most common way beginners get an unpleasant bill.
Official source: Anthropic developer documentation
Create an account on the developer console — this is separate from any chat subscription you may already pay for, and one does not include the other.
Generate an API key and treat it like a password. Anyone holding it can spend your money.
Store the key in an environment variable, never inside your code. Keys committed to a public repository are found and abused within minutes.
Set an organisation spend limit before writing anything that runs in a loop.
Start with the cheapest model that could plausibly work. Moving up is easy; discovering you overpaid for a month is not.
This is the same company as the Claude chat assistant and the default is reversed. API traffic falls under commercial terms, which means it is not used to train models by default, and inputs and outputs are deleted from backend systems within about thirty days. The consumer chat product trains on personal-plan conversations unless you opt out; the API does not train on them at all.
The transferable lesson: which door you walk through determines which contract you are under. Two interfaces to the same underlying model can have opposite data defaults, and neither one advertises this at the moment you paste something sensitive.
Uses the console's testing workbench to send one message to a model and read the reply, with no program involved.
Why this one: Confirms your key, your account, and your billing work, without a debugging session in the way.
Tweak: Send something deliberately trivial. You are testing the plumbing, not the model.
Runs one representative item, then multiplies by your real volume.
Why this one: Per-call costs look negligible and per-month costs frequently do not. Do this arithmetic while it is still cheap to change your mind.
Tweak: Price it against the cheapest model too. The difference is often larger than the quality gap.
Submits a set of requests together for processing later at half the per-token rate, instead of sending each one and waiting for its answer.
Why this one: Half price in both directions is the biggest lever available, and it costs nothing but patience.
Tweak: If you find yourself waiting on results interactively, the job was not a batch job.
Stores the key in your system's settings, where your program reads it at run time, so it never appears in the code itself.
Why this one: Every leaked-key story starts with "I was going to fix that later".
Tweak: Add the environment file to your ignore list in the same commit, not the next one.
Want the full picture first? See the Anthropic API overview for pricing, the free tier and related tools.
