How to Get a Claude API Key (and What It Costs) in 2026

Section 01What the Claude API is and how it differs from a subscription

Claude has two doors. The first one everybody knows: the claude.ai site, the phone app, the chat window. You type, Claude answers. You pay with a subscription: Pro at $20 a month or Max at $100 and $200.

The second door is the API. It is the entrance for programs. Your Telegram bot receives a client's question, forwards it to Claude by itself, gets the answer, and sends it back to the client. No human in the chain. A site with an assistant, a spreadsheet that sorts incoming leads into categories on its own, an agent that builds a sales report every morning – all of that runs through the API.

The one thing to remember from the start: these are two separate wallets. A Pro subscription gives you zero free API requests. An API balance does not unlock Pro in the chat. Even the accounts live in different places: the chat is at claude.ai, the API is in the developer Console at platform.claude.com. You can sign in with the same email, but each door has its own billing.

How the money is counted

A subscription works like a gym membership: you pay a fixed amount and use it within the limits. The API works like an electricity meter. You put money on the balance up front, and every request takes its share depending on how much text went into Claude and how much came back.

Text is measured in tokens. A token is a piece of a word. The official models page gives a conversion: a million tokens is roughly 555,000 English words. For quick math, think of a dense page of text as about a thousand tokens. Prices are listed per million tokens, separately for input (what you sent) and output (what Claude wrote). Output always costs five times more than input.

What you can build with the API

• A Telegram bot that answers clients in your words – how to build one with no programmer is covered in Telegram bot without a programmer.
• An assistant on your site that knows your programs and prices. Example: an AI consultant on your site.
• An agent in your DMs that replies to leads and books calls: an AI agent in your Telegram DMs.
• Bulk processing: sort 500 reviews, rewrite 200 product descriptions, check 50 homework assignments overnight.
• Connecting Claude to no-code services: Make, n8n, and similar builders ask for exactly this API key.

If you want Claude for yourself in a chat window, start with Claude from scratch: sign-up and plans are covered there. This article is about the next step, when Claude has to work inside your services on its own.

Section 02What the Claude API costs in 2026

Haiku 5.5 has two-tier pricing. A request up to 100,000 tokens costs $0.10 for input and $0.50 for output. Above that, the price rises to $0.50 and $2.50. A bot with short questions almost always stays in the lower tier.

All four models have a context window of one million tokens and can write up to 128,000 tokens in a single answer. Their knowledge runs through June 2026.

Discounts worth knowing

• Batch processing. If you don't need the answer instantly, send tasks as a batch through the Batch API: the price drops by 50%. Good for overnight processing of reviews, descriptions, and homework.
• Prompt caching. If every request repeats the same large chunk, say a ten-page bot instruction, you can cache it. Reading from the cache on Sonnet 5.5 costs $0.10 instead of $2 per million tokens, on Opus 5.5 $0.20 instead of $4.
• Fast mode. Opus 5.5 also has the opposite option: answers 2.5x faster at twice the price. A beginner doesn't need it.

What that means in real life

"Dollars per million" says nothing on its own, so here is a live example. A therapist's bot answers 1,000 client questions a month. Each request carries the bot instruction plus the question – about 1,500 input tokens. The answer is about 300 tokens. Over a month that comes to 1.5 million input tokens and 0.3 million output tokens.

A thousand client replies for 30 cents. Really. That's why the $5 minimum top-up lasts a beginner a long time. Money starts to melt when you run Opus or Fable on huge documents or forget to stop a script stuck in a loop. How to protect yourself from that is in the section on saving money.

Is it free?

No. The official help center says it plainly: the API and the Console's Workbench are paid with prepaid credits. As of October 31, 2026, Anthropic does not promise any free trial balance. If someone offers you a "free Claude key," it is either somebody's stolen key or a reseller who will send you a bill later. Use neither.

Section 03Signing up in the Console: step by step

The developer Console is your API account. It holds your balance, your keys, your spending stats, and the Workbench, where you can talk to a model without writing code. The API works in a long list of countries, the full list is on the supported regions page.

Section 04How to buy credits

With an empty balance the key won't work: the very first request returns an error about insufficient credits. So payment comes first, the key second.

Section 05How to create an API key

The key is a password that tells Anthropic the request came from you and which balance to charge. It takes a minute to create.

Section 06Your first request: with and without code

Option 1. No code, right in the browser

The Console has the Workbench, a sandbox. Pick a model, write a system instruction and a question, click Run. It's the same API, only with buttons. This is the best place to debug a bot instruction before you hook it up to Telegram. Keep in mind that Workbench requests are charged to your balance too.

Option 2. Let Claude Code do everything for you

If you already have Claude Code installed (how to do that is in how to use Claude Code), you don't need to set anything up by hand. Tell it:

I have a Claude API key. Store it somewhere safe on this computer so it never ends up in the code or in git, then make a test request to the claude-sonnet-5-5 model with the question "Answer in one word: are you working?". Show me the answer and how many tokens it used.

Claude Code creates the files it needs, asks for the key, and shows you the result. Your only job is to paste the key when it asks.

Option 3. One command in the terminal

For those who aren't afraid of the terminal. First save the key to an environment variable:

export ANTHROPIC_API_KEY="sk-ant-your-key"

Then send a request. This is the official example from the docs, I only changed the model and the question:

curl https://api.anthropic.com/v1/messages \
  -H "content-type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-sonnet-5-5",
    "max_tokens": 500,
    "messages": [
      {"role": "user", "content": "Hi! Answer in one sentence: are you working?"}
    ]
  }'

You get JSON back. Claude's text sits in the content field, and the usage field shows how many tokens went in and came out. With those two numbers you can always work out what a request cost.

Option 4. Python

pip install anthropic

import anthropic

client = anthropic.Anthropic()  # the key is read from ANTHROPIC_API_KEY

message = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=500,
    system="You are a therapist's assistant. Answer warmly and briefly.",
    messages=[{"role": "user", "content": "How long is a session?"}],
)

print(message.content[0].text)
print(message.usage)

Model IDs for the API as of October 31, 2026: claude-haiku-5-5, claude-sonnet-5-5, claude-opus-5-5, claude-fable-5-1. Not sure where to start? Start with Sonnet 5.5: it's fast and smart. Move to Haiku 5.5 later for simple bulk tasks, once you see Sonnet handles them with room to spare.

Section 07A live test

The most common reason people get an API key is a bot that answers clients. Its heart is the system instruction. Here's a working instruction for a nutritionist's bot. Swap in your own details and debug it in the Workbench.

You are the assistant of Anna, a nutritionist. You answer people on Telegram who are interested in her program "Easy Month."

Program facts:
– 4 weeks, a new group starts every Monday
– a menu for every day plus a shopping list
– a chat with Anna, replies within the day on weekdays
– price $129, pay by card or in installments
– not suitable during pregnancy or for people with diabetes; in those cases, a personal consultation only

Rules:
– Keep answers short, 2–4 sentences, polite and warm.
– Never diagnose and never give medical advice.
– If the question isn't covered by the facts, say you'll check with Anna and ask for a phone number.
– At the end of the answer, when it fits, offer to book a free 15-minute call.

What did that conversation cost? The instruction plus the question is about 500 tokens, the answer about 150. On Sonnet 5.5 that's roughly a quarter of a cent. On Haiku 5.5, hundredths of a cent. A thousand such conversations on Sonnet fit into $2–3.

A tip from practice: before launch, throw 20 awkward questions at the instruction. About discounts you don't offer. About contraindications. About competitors. A rude question. An off-topic question. Every answer you don't like becomes a new rule in the instruction. That way the bot gets more reliable in one evening than in a week of mistakes in front of real clients.

Section 08Where to run it 24/7

While you test in the Console, everything runs from your browser. A bot that answers clients around the clock can't live on your laptop, though. Close the lid and the bot goes quiet. Wi-Fi drops at night and so does the bot.

The right home is a small server: a computer in a data center that runs nonstop. Any small VPS or cloud server will do. The bot lives there and sends its requests to Claude day and night. For one or two bots the most basic plan is plenty: 1–2 CPU cores, 2 GB of RAM, a 20 GB disk. Pick a data center location in a supported region, and you're set.

You don't have to configure the server by hand. Install Claude Code on it and then just talk: "deploy a Telegram bot here, here's the Claude API key, here's the bot instruction, make it restart itself after a crash." If Claude Code is new to you, start with Claude Code for non-coders.

One more path with no server at all: automation builders like Make or n8n in the cloud. You paste the key there, and the workflow calls Claude on its own. The downside is a subscription to the builder itself and less flexibility.

Section 09How not to overspend

Across 250+ projects rolling out AI for therapists, coaches, lawyers, and marketplace sellers, I've seen money leak out of an API balance through the same few holes. Here's what closes almost all of them.

1. Set a spend limit on day one

Settings, Billing, the Spend limits section, the Set limit button. Enter the most you're willing to spend in a month, say $10. Once spending reaches it, the API starts returning an error and stops charging you. This is the main insurance against a looping script that fires ten thousand requests overnight.

Anthropic also has its own ceiling based on your account tier. On the Start tier it's $500 a month, on Build it's $1,000. New accounts sometimes begin on an even lower Evaluation tier, and the limits there go up by themselves as you use the API.

2. Match the model to the task

Using Fable 5.1 to sort leads is like taking a limo to buy bread. The price gap between Haiku and Fable is a hundredfold. The rule is simple: start with Sonnet 5.5, then try the same task on Haiku 5.5. If the quality holds up, switch. Bring in Opus and Fable only where Sonnet actually makes mistakes.

3. Cap the answer length

The max_tokens parameter in a request is the ceiling on answer length. For a bot, 300–500 is enough. Output costs five times more than input, so a single line in the instruction, "answer in 2–4 sentences," saves money directly.

4. Don't drag the whole history into every request

The API doesn't remember past messages by itself. For a bot to remember the conversation, the program resends the entire thread every time. By the tenth message you're paying for all ten. Keep the last 10–15 turns and compress older ones into a short summary.

5. Cache the long instruction

If the bot instruction is long (price list, programs, answers to common questions), turn on caching. A cache read costs 5% of the normal input price on Sonnet and Opus 5.5. Tell Claude Code: "turn on prompt caching for the bot's system instruction."

6. Bulk jobs go in a batch

Sorting 300 reviews by morning is a job for the Batch API. Half the price, results within 24 hours, usually much sooner.

7. Check Usage once a week

The Usage page shows spending by day and by model. A sudden spike is your cue to check whether a script got stuck or a key leaked.

Section 10How to keep the key safe

The key is access to your money. Whoever has it spends your balance. Bots crawl the internet around the clock looking for sk-ant- keys in public code, and they find them within minutes.

• Don't send the key in messengers, even to yourself in Saved Messages. Don't put it in screenshots or shared documents.
• Don't write the key into your code. It belongs in an environment variable or a secrets file that never goes into git. Claude Code will set that up if you ask.
• One key per project. If the site's key leaks, you disable just that one and the bot keeps working.
• Give a contractor a separate key with a short lifetime, and delete it when the work is done.
• Suspect a leak? Delete the key on the API keys page right away and create a new one. Then check Usage for spending that isn't yours.

More on protecting your projects from AI-powered attacks is in how to protect your blog and client base.

Section 11When something goes wrong

The fastest way to deal with any error is to copy its full text and hand it to Claude Code or to Claude in the chat with the question "what does this mean and how do I fix it." Every API response includes a request_id. If it ever comes to Anthropic support, send it along with your description.

Section 12API, subscription, or Claude Code: which to choose

In practice all three work together. The subscription is for your daily work. Claude Code is for building the bot or agent. The API is what keeps that bot running on its own afterwards. If you're still choosing an AI, the comparison with its main rival is in Claude vs ChatGPT. And to connect Claude to Telegram, spreadsheets, and your site without programming, see MCP in Claude.

Section 13The whole path in a minute

Start today with one action: top up $5 and run the bot instruction from section 7 in the Workbench against ten real questions from your clients. Five dollars, and you'll know for sure whether AI can handle your conversations.