Claude API free tier? No. What a first app costs

Does the Claude API have a free tier?

The Claude API has no free tier. There is no monthly allowance of free tokens, and no plan you can sit on for $0. What a new account does get is a small amount of free credits to test the API, and Anthropic's pricing docs do not state the amount. Once those credits are spent, requests draw on paid credit or other API credit grants. If you also pay for a Max, Ultra or Team subscription, those plans include a monthly API credit, covered below. And if you research AI safety or alignment, Anthropic's External Researcher Access Program applies $1,000 in API credits to the account of each approved applicant, as of October 2026. Anthropic reviews submissions on the first Monday of each month and does not reply to applications it turns down.

That is the direct answer. The rest of this guide is what follows from it: which parts of the platform cost nothing at all, why this is a different question from the free plan in the Claude apps, and the cheapest honest way to get a first program working.

What is free on a Claude API account

Several pieces of the platform carry no token charge, and they matter more than they sound.

• Counting tokens. The token counting endpoint is free to use. It has its own requests-per-minute limit, separate from message creation, so measuring a prompt never eats into the budget for sending it.
• The Console itself. Creating an organization, reading your usage page, and opening the playground (formerly the Workbench) cost nothing. Running a prompt in the playground is an ordinary API request, so that part bills tokens like any other call.
• File operations. Uploading, listing and deleting files through the Files API is free. You pay only when a file's content enters a request as input tokens.
• Web fetch. The web fetch tool adds no charge beyond the tokens of whatever it pulls into the conversation. Web search is not free: it is $10 per 1,000 searches as of October 2026.
• Code execution, up to a point. Each organization gets 1,550 free container hours per month as of October 2026, then $0.05 per hour per container. Code execution costs nothing extra in a request that includes web search or web fetch in their 20260209 versions or later.

None of that is a free tier. It means the metering is honest: you are billed for inference, and not for the plumbing around it. Free web fetch pairs well with data sources that are themselves free, and running stock and options analysis off a keyless market data API is a worked example where the only thing you pay for is the tokens the fetched data brings into the prompt.

Price a prompt before you pay for it

Because counting is free, you can learn the size of a request first. The count is an estimate, and Anthropic's docs say the tokens a real request uses can differ by a small amount. This is the same endpoint the SDKs wrap, and it takes the same body as a real message, with a few exceptions: it rejects requests that carry server tools such as web search, web fetch or code execution, the MCP connector, or images and PDFs passed by URL or file ID, so send images and PDFs as base64 to count them.

Use a key scoped to a workspace for the example below. If your key is not scoped to a workspace, also send anthropic-workspace-id with the workspace's ID; omitting it returns HTTP 400 with invalid_request_error.

curl https://api.anthropic.com/v1/messages/count_tokens \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "content-type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "claude-haiku-4-5",
    "system": "You are a scientist",
    "messages": [{"role": "user", "content": "Hello, Claude"}]
  }'

A healthy response is one line of JSON, such as {"input_tokens": 14}. Multiply that by your model's input rate and you have a close estimate of the request's input cost before you send it. Count against the model you will actually call, because Claude 4.7 and later models produce roughly 30% more tokens for the same text than older ones such as Haiku 4.5. If you get HTTP 401 with authentication_error, check whether the key is wrong, revoked or expired. A correctly supplied key also returns this error after it expires; check its expiration in the Console and create a new key, because expired keys cannot be reactivated. If the key looks right and the call still will not authenticate, or you are reaching Claude through Bedrock, Vertex or Foundry rather than an Anthropic key, the four ways to authenticate a client sets out what each one expects and where to keep the credential.

Why this is a different question from the free plan in the apps

People search for both, and the two have almost nothing to do with each other. The free plan on the Claude apps is a real product you can use for a long time without a card, metered by a rolling usage window, and what the Claude free tier actually includes covers what it carries and when it resets. The API is prepaid metering with no such plan behind it.

The consequence that trips people up is that on most plans a subscription and an API key are separate wallets. Paying for Pro does not fund a key, and API credits do not raise your chat allowance. Claude Code is where that split confuses people most, because it runs on either one: whether Claude Code is included with Pro sets out which plans cover it and when it quietly falls back to billing a key. If you are choosing between the two rather than trying to avoid paying at all, Claude API cost versus a subscription works through the arithmetic on real workloads.

There are two exceptions. Since 7 October 2026, Max and Team plans include a monthly API credit. As of October 2026 it is $100 on Max and $200 on Ultra, formerly Max 20x. On Team it is $20 per Standard seat or $100 per Premium seat, pooled and capped at $500. You claim it by linking a Console organization (new subscribers wait 7 days first), and you do not need to add a payment method to the Claude Platform. It covers the API, the Agent SDK and Managed Agents but not interactive Claude Code. It also covers claude -p when you run it yourself with an API key from the linked Console organization; runs started by the Claude Code GitHub Action, an IDE extension or the Claude desktop app are not covered. Unused credit expires at the end of each billing cycle. Once it is spent, requests stop until the next credit arrives unless you have also bought API credit or have an invoicing arrangement. Free, Pro and Enterprise plans do not get it, so read what to weigh before upgrading to Max before you buy a plan mainly for the credit. The other exception is Claude Enterprise, which pairs a per-seat price with token usage metered at standard API rates, and how Enterprise seat pricing works covers what the seat minimum buys before the meter starts.

What a request costs once the credits are gone

These are the first-party API rates as of October 2026, per million tokens.

[
  {
    "label": "Haiku 4.5",
    "input_usd": 1,
    "output_usd": 5
  },
  {
    "label": "Sonnet 5",
    "input_usd": 2,
    "output_usd": 10
  },
  {
    "label": "Opus 5",
    "input_usd": 5,
    "output_usd": 25
  }
]

Haiku 4.5 bills 1 dollar per million input tokens and 5 dollars per million output tokens. Opus 5 bills 5 and 25 for the same volumes, so the model you pick moves a bill by 5x before any prompt tuning does. Sonnet 5 launched at $2 and $10 as an introductory rate, and Anthropic has since made that its standard price, so the increase to $3 and $15 that was scheduled for 1 September 2026 did not happen. Output costs five times what input costs on every one of them, which is the most useful fact on this page: a chatty answer is more expensive than a long question. which Claude model you should use compares them on a worked job. Fable 5 sits above all three at $10 per million input tokens and $50 per million output, so what those rates actually buy is worth settling before you point anything routine at it. Three newer models are also on the price list: Haiku 5.5 bills $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, Sonnet 5.5 bills $2 and $10, and Opus 5.5 bills $4 and $20.

The cheapest real paths to experiment

Four levers do almost all the work, and none of them needs a discount.

• Start on the smallest model that passes your test. Classification, extraction, tagging and routing rarely need a frontier model. Haiku 4.5 is a fifth the price of Opus 5 and answers faster.
• Cap max_tokens. It is a hard ceiling on generated tokens, and output is the expensive half. A classifier that answers in one word should not carry a 16,000-token budget.
• Cache the part of the prompt that never changes. A cache read costs 10% of the base input price on most models (5% on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1) and a five-minute cache write costs 1.25x, so caching pays for itself after a single read. Caching is a prefix match, which means the stable text has to come first and the varying question last, or nothing matches and you pay the write premium for nothing.
• Batch anything that is not interactive. The Message Batches API is 50% off both input and output, and most batches finish inside an hour. Nightly jobs, backfills and evaluation runs belong there.

A worked cost for a first small app

Take a log-summary script: 1,000 calls, each sending about 800 input tokens of log lines and getting back about 200 tokens of summary. That is 800,000 input tokens and 200,000 output tokens on Haiku 4.5.

[
  {
    "label": "Standard rates",
    "cost_usd": "1.80"
  },
  {
    "label": "Same job on the Batch API",
    "cost_usd": "0.90"
  }
]

The arithmetic is 0.8 million input tokens at $1 plus 0.2 million output tokens at $5, which is 1.80 dollars in total. Sent as a batch instead, the same work costs 0.90. A starter credit balance covers a lot of testing at that scale, which is why the absence of a free tier bites much later than people expect. For the full walkthrough on a server you own, building a first Claude API app on a VPS covers key handling, streaming and the error paths.

Usage tiers are ceilings, not allowances

The API places every organization on a usage tier automatically, based on usage history and account standing. The tiers are named Start, Build, Scale and Custom, and each one sets rate limits plus a monthly spend cap. As of October 2026 the caps are $500 on Start, $1,000 on Build and $200,000 on Scale, and Custom has no cap. New organizations, and those with little usage history, may start on an Evaluation tier with limits below these until that history builds.

Read that as a limit rather than a grant. Being on the Start tier does not mean you have $500 to spend. It means you may not spend more than $500 in a calendar month before usage pauses until the next month. You can also set your own lower spend limit under the tier cap, which is the first thing to do on any account attached to a card, because a runaway loop is a far more common way to lose money than the per-token price is.

What to do if free is a hard requirement

If the budget is genuinely zero and has to stay there, the API is the wrong tool, and no amount of tier shopping changes that. Two honest options remain. Use the free plan in the Claude apps for interactive work, accepting that it is metered by window and gives you no programmable endpoint. Or run an open-weights model on hardware you rent, which swaps a per-token bill for a server bill you control: running Ollama on a VPS to self-host an LLM covers the memory each model size needs and what the small ones can actually do. If it turns out the usage window is the real constraint rather than the money, Claude Pro at $20 a month lifts that ceiling in the apps for less than many first API bills. It comes with no API key and no API credit, though. The closest it gets to calling Claude from code is Claude Code, which Pro includes and which can run non-interactively with claude -p against the same usage limits. Whether that $20 is worth paying comes down to how often the free limit actually stops you, which a decision guide for Pro works through against real usage patterns. If you are shopping across providers rather than within one lineup, Claude and ChatGPT plans priced against each other sets the free and paid tiers on both sides together, including which comes out cheaper for coding work. And if a subscription you are already paying for is part of the spend you are trying to zero out, cancelling or downgrading that plan keeps the month you have already bought and stops the next one.

For anything in between, the practical answer is a funded key, a low spend limit, a small model, and prompt caching turned on from the first day.

FAQ

Is the Claude API free to use?

No. There is no free tier and no monthly free token allowance. New accounts receive a small amount of free credits to test the API. After that, requests draw on paid credit or other API credit grants, including the monthly API credit included with Max, Ultra and Team plans, which needs no payment method on the Claude Platform, and the $1,000 in API credits that Anthropic's External Researcher Access Program applies to the accounts of approved AI safety and alignment researchers. Some parts of the platform do cost nothing, including token counting, file uploads and deletions, and the web fetch tool beyond the tokens of what it fetches.

Do I get free credits when I sign up for the Claude API?

Yes, a small amount, and Anthropic's pricing docs do not give the figure. Treat it as enough to prove your integration works rather than as a budget to build on. Extended trial credit for enterprise evaluation is arranged through Anthropic sales, not granted automatically.

Does paying for Claude Pro or Max include API access?

Pro does not, and Max now includes some. A Pro subscription and an API key are billed separately and metered differently: Pro raises your allowance inside the Claude apps and Claude Code, and nothing more. As of October 2026, Max includes a monthly API credit of $100 and Ultra includes $200, which you claim by linking a Console organization. It covers API calls but not interactive Claude Code, expires each billing cycle, and does not change your limits in the apps. Past that credit, an API key draws on prepaid credits and is limited per minute by requests and by tokens.

What is the cheapest way to test the Claude API?

Count your tokens first, since that endpoint is free. Then run the test on Haiku 4.5 with a tight max_tokens, cache a fixed prompt prefix of at least 4,096 tokens so repeat calls read at 10% of the input price, and send anything that is not interactive through the Batch API at half price. On those settings a thousand small calls costs under a dollar as of October 2026.