Tokens to Characters: How Many Characters Is a Token?

The 4-characters-per-token rule

For English prose, dividing the character count by four lands within about 20% of the true token count for most real text. That makes it the standard back-of-napkin converter: a 2,000-character email is roughly 500 tokens; a 40,000-character document is about 10,000. The range to plan against is 3–5 characters per token.

The rule exists because BPE tokenizers give frequent English patterns their own tokens. The word understanding might be two tokens; the, and, and is are one each. Punctuation usually earns its own token, which is why dialogue-heavy and formulaic text runs denser than average.

Character-to-token conversion table

Translate common character counts using the 4-character rule with honest ranges:

• 100 characters ≈ 25 tokens (20–33)
• 250 characters ≈ 63 tokens (50–83)
• 500 characters ≈ 125 tokens (100–167)
• 1,000 characters ≈ 250 tokens (200–333)
• 2,000 characters ≈ 500 tokens (400–667)
• 5,000 characters ≈ 1,250 tokens (1,000–1,667)
• 10,000 characters ≈ 2,500 tokens (2,000–3,333)
• 50,000 characters ≈ 12,500 tokens — a short paper
• 500,000 characters ≈ 125,000 tokens — a full book

Where the rule breaks

Source code is the biggest exception in practice: indentation, operators, brackets, and identifiers push density toward 3 characters per token or worse. Numbers and IDs split unpredictably — a long hash can cost many tokens. URLs and emails fare badly too, since they contain rare character sequences.

Non-Latin scripts change the ratio entirely. CJK text (Chinese, Japanese, Korean) often costs 1–1.5 characters per token, meaning 1,000 Japanese characters can be 700+ tokens instead of 250. Languages in Cyrillic or Arabic script sit between the extremes. If your product accepts international input, measure per language rather than trusting ÷4.

From characters to cost

Character counts are usually what you have — a document, a paste, a database field — while pricing is per token. The chain is: characters ÷ 4 gives estimated tokens; tokens ÷ 1,000,000 × the per-million price gives cost. A 120,000-character report is ~30,000 tokens, costing about $0.06 as input to Claude Sonnet 5 — and about $0.30 if the model generates 30,000 tokens back at $10 per million.

For anything that matters, replace the estimate with a real count. PromptCostLab's converter and prompt calculator run the o200k tokenizer client-side: exact for OpenAI-family models, a labeled approximation for Claude, Gemini, Grok, Kimi, and DeepSeek.

Why providers don't publish one ratio

Every model family trains its own tokenizer vocabulary, so the same characters tokenize differently across GPT, Claude, Gemini, and the rest — and tokenizer generations change within a provider (Anthropic's 2026 refresh increased tokens for identical text by roughly 30%). A ratio is a planning tool, not a physical constant. Providers that expose counting endpoints (Anthropic's count-tokens, Gemini's countTokens) are the authoritative pre-flight check when the number must be right.