What Is Anthropic API Tokens? A Clear Explanation

What Anthropic API Tokens Actually Are

Tokens are the unit the Anthropic API uses to measure text. Every request you send to Claude — the system prompt, the conversation history, the user message — gets broken into tokens before the model processes it, and every response Claude generates is also counted in tokens. Tokens are not words and not characters; they're chunks of text produced by a tokenizer, and that distinction matters because it's what determines both your bill and the limits on how much text you can send or receive in a single call.

If you're asking "what is Anthropic API tokens," you're probably trying to understand one of two things: how pricing works, or why a request got cut off or rejected for being too long. Both come down to the same concept. Anthropic (and any provider reselling Claude access, including SubToAPI) prices usage per token and enforces a maximum context window measured in tokens. Understanding how tokens are counted is the fastest way to predict cost and avoid hitting limits unexpectedly.

How Text Gets Turned Into Tokens

Claude's tokenizer splits text into subword pieces, not whole words. A rough rule of thumb for English text:

• 1 token ≈ 4 characters of typical English text
• 1 token ≈ ¾ of a word on average
• 100 tokens ≈ 75 words

So a 500-word blog paragraph is roughly 650–700 tokens. Code, non-English languages, and text with lots of punctuation or unusual formatting tokenize less predictably — code especially tends to use more tokens per character than prose because of symbols, indentation, and short identifiers.

This is why token counts aren't intuitive from word count alone. A JSON payload with deeply nested keys can cost more tokens than a plain sentence of similar length.

Input Tokens vs Output Tokens

Every API call has two token counts that matter:

• Input tokens: everything sent to the model — system prompt, message history, tool definitions, and any documents or context you include
• Output tokens: everything the model generates in its response

These are usually priced differently, with output tokens typically costing more per token than input tokens. This matters for how you design prompts: a long system prompt reused across thousands of requests adds up in input tokens even if each individual response is short. Conversely, asking the model to generate long structured output (detailed JSON, long-form text) drives up output token cost regardless of how short your prompt was.

Where Token Counts Show Up in a Response

When you call the API, the response includes usage metadata showing exactly how many input and output tokens were consumed. Here's what that looks like with SubToAPI, which sits on top of your existing Claude access and returns the same kind of usage data:

curl https://api.subtoapi.app/v1/messages \
  -H "Authorization: Bearer $SUBTOAPI_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-4-5",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "Summarize the key points of REST API design."}
    ]
  }'

The response includes a usage object with input_tokens and output_tokens. Logging this on every call is the simplest way to track cost per feature, per user, or per endpoint without guessing. See /docs/messages for the full request and response shape.

Why Token Limits Exist

Every Claude model has a context window — a maximum number of tokens it can consider at once, combining input and output. If your conversation history plus system prompt plus new message exceeds that window, the request fails or gets truncated depending on how your client handles it. This is separate from max_tokens, which is a parameter you set to cap how many tokens the model is allowed to generate in its response — useful for controlling cost and latency on a per-request basis.

Long-running chat applications need to manage this actively: trimming old messages, summarizing history, or using shorter system prompts as conversations grow. Streaming doesn't change token counting — it changes how the response is delivered, not how much you're charged. Details on streaming behavior are in /docs/streaming.

Estimating Tokens Before You Send a Request

You generally don't need exact token counts before sending a request — the character-based rule of thumb (roughly 4 characters per token for English) is close enough for budgeting and setting max_tokens sensibly. For production systems where cost predictability matters more, track actual usage data returned from each call rather than estimating upfront. Over time this gives you a much more accurate picture than any static conversion rate, especially if your traffic mixes plain prose, code, and structured data.

Tokens and Cost Tracking in Practice

If you're building a product on top of Claude, token usage is the number that determines your margins. A support bot with long system prompts and short answers has a very different cost profile than a code generation tool with short prompts and long outputs. Reviewing input vs output token ratios across your actual traffic — not assumptions — is the only reliable way to price your own product or decide where to trim prompts.

SubToAPI exposes this usage data per request and rolls it up at the team level, so you can see token consumption across API keys and seats in one dashboard instead of piecing it together from logs. If you're already paying for Claude access and want that turned into a standard HTTPS API with usage visibility built in, /docs/quickstart walks through getting a key and making your first call.

FAQ

Are Anthropic API tokens the same as words? No. A token is roughly ¾ of an English word on average, but the exact ratio depends on the text — code, symbols, and non-English languages tokenize differently than plain prose.

Do I pay for both input and output tokens? Yes. Input tokens (your prompt and context) and output tokens (the model's response) are counted and typically priced separately, with output tokens usually costing more per token.

How do I see how many tokens a request used? The API response includes a usage object with exact input_tokens and output_tokens counts for that call — see /docs/messages for the response format.