Claude token counter
Estimate how many tokens your text or conversation will use on Anthropic’s Claude models — Opus, Sonnet, and Haiku. Token counts drive both billing and the 200K context window, so checking before you send keeps long documents and multi-turn chats inside the limit.
Why token counts matter
Claude’s API bills by the token — separately for input (prompt) tokens and output (completion) tokens. Token counts also determine whether a request fits within the model’s context window. Claude models support a 200K token context window, which is large but not unlimited. A long document, a full codebase pasted as context, or a lengthy multi-turn conversation history can approach or exceed that limit, causing the API to return an error or silently truncate earlier context depending on how the request is structured.
Checking the token count before sending helps you:
- Stay within the context window — confirm that your system prompt + full conversation history + documents + user message all fit in one request.
- Estimate costs before a batch run — if you’re processing thousands of documents, knowing the average tokens per document lets you project spend before starting.
- Optimize prompts for efficiency — identifying unexpectedly token-heavy sections of a prompt helps you trim without losing meaning.
How the estimate works
Claude uses its own byte-pair encoding vocabulary, distinct from OpenAI’s tiktoken. This tool applies Claude’s measured average of roughly 3.7 characters per token for English, blended with a word-boundary heuristic. When you paste a JSON array of messages, each message is counted separately and a small fixed overhead per message approximates the role and structural tokens the Messages API adds around each turn.
Understanding the multi-turn conversation cost
One aspect of Claude’s billing that surprises many users: every turn in a conversation resends the full history. In a standard stateless API call, you include the entire prior conversation in the messages array. Turn 1 sends 1 message; Turn 5 sends all 5 messages. This means token consumption grows quadratically with conversation length in the worst case.
For example (illustrative only):
If each user message and assistant reply averages 200 tokens:
| Turn | Messages sent | Total input tokens that turn |
|---|---|---|
| 1 | 1 | ~200 |
| 5 | 5 | ~1,000 |
| 10 | 10 | ~2,000 |
| 50 | 50 | ~10,000 |
For a long-running assistant application, conversation management (summarizing old turns, truncating early history) becomes a cost and context-window concern simultaneously.
Claude’s tokenizer vs GPT tokenizers
Claude uses a different byte-pair encoding vocabulary from OpenAI’s tiktoken. In practice:
- For typical English prose, Claude tends to use slightly more tokens per character than GPT-4o — roughly 3.6–3.8 characters per Claude token vs approximately 4 characters per GPT-4o token.
- Code tends to tokenize more efficiently in both models but the ratio still differs.
- Non-English text, especially languages with complex scripts, varies considerably by tokenizer design.
The practical implication: a GPT-4o token estimate will undercount Claude tokens by a small but real amount. Do not use a tiktoken count as a budget proxy for Claude without applying a correction factor.
Tips and notes
- Claude usually counts slightly more tokens than GPT-4o for identical English text, so don’t reuse a GPT estimate as-is.
- A multi-turn conversation accumulates: every prior turn you resend counts again toward the context window and the bill.
- System prompts count as input tokens on every API call, even though the system prompt content rarely changes between calls.
- For exact counts in production, use Anthropic’s
count_tokensendpoint, which returns the precise token total for a given request. This tool provides a useful fast estimate; the API endpoint provides the authoritative count.