Claude token counter

How the estimate works, and how far to trust it

The page splits your text the way tokenizers tend to: common words whole, long words in pieces, each Chinese or Japanese character on its own, punctuation apart. It then scales the result for the model. It's fitted against a public tokenizer, not against Claude's, which isn't published. The range is our judgement, not a measured error bar. Treat it as good enough for “will this fit?” and “roughly what will it cost?”, and expect it to be furthest off on code, non-English text and emoji.

The estimate covers your text alone. A real request adds a little: Anthropic's documented example of a short system prompt plus Hello, Claude counts as 14 tokens. Tool definitions add more, and thinking blocks from the current turn count too.

Count Claude tokens in code

For anything you'll act on, use the count_tokens endpoint with the model you plan to run. It takes the same input as a real message, including system prompts, tools, images and PDFs:

curl https://api.anthropic.com/v1/messages/count_tokens \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5-5",
    "system": "You are a scientist",
    "messages": [{"role": "user", "content": "Hello, Claude"}]
  }'
# {"input_tokens": 14}

import anthropic

client = anthropic.Anthropic()
response = client.messages.count_tokens(
    model="claude-opus-5-5",
    messages=[{"role": "user", "content": open("CLAUDE.md").read()}],
)
print(response.input_tokens)

It's free, and it's rate limited on its own: 5,000 requests a minute on the Start usage tier at the time of writing. Recount when you change models: a count from before Claude Opus 4.7 won't carry over.

What the cost column means

• Input only, one request. At each model's standard API price. Output usually costs five times as much per token, and it depends on the answer, which nobody can count in advance.
• A long conversation resends its history. Each turn pays for everything before it again, unless prompt caching covers it, which makes a big system prompt much cheaper than this column shows.
• Claude Haiku 5.5's rate changes with size. $0.10 per million tokens up to 100,000 tokens, $0.50 per million for longer prompts.

For how these prices have moved, see the price of AI tokens, 2020–2026 and AI token spending in 2026.

Frequently asked questions