2026 edition · companion to the free token counter
A token is the unit of text that a language model reads and writes. Before your prompt reaches the model, a tokenizer splits it into these chunks using a fixed vocabulary of sub-words learned from training data. The word cat is one token; tokenization might split into token + ization; a long UUID or a chunk of base64 splits into many tokens because those byte sequences are rare in the training corpus.
Rules of thumb for English: 1 token ≈ 4 characters ≈ 0.75 words, so 1,000 tokens ≈ 750 words ≈ 3 pages of prose. For Chinese, Japanese and Korean, expect roughly 1 token per character — a 1,000-character Chinese document costs more than double its English translation.
There is no universal tokenizer. As of late 2026:
| Model family | Tokenizer | Count accuracy |
|---|---|---|
| GPT-4o / GPT-4.1 / o-series | o200k_base | Exact counts available (e.g. our counter) |
| GPT-4 / GPT-3.5 | cl100k_base | Exact counts available |
| Claude (Anthropic) | proprietary | Estimate only (~4 chars/token) |
| Gemini (Google) | proprietary (SentencePiece-based) | Estimate only |
If your budget depends on precision, count with the exact encoder of the model you're calling — that's what this site does for the OpenAI families.
Suppose your application sends a 2,000-token system prompt with every request, and users average 500-token messages, with 300-token replies coming back. Each request burns 2,500 input tokens and 300 output tokens. At GPT-4o list prices ($2.50 per 1M input, $10 per 1M output):
input: 2,500 × $2.50 / 1M = $0.00625 · output: 300 × $10 / 1M = $0.003 → about $0.0093 per request.
Now the leverage: 10,000 requests a month cost ~$93 — but trim the system prompt to 800 tokens and you cut input cost by more than half, landing near $50. This is why counting tokens before shipping is not pedantry; it's margin engineering.
1. Trim and cache the system prompt. Delete instructions you never observe being followed anyway; enable provider-side prompt caching so the repeated prefix is billed at a discount.
2. Summarize chat history. Rolling windows or periodic summaries keep long conversations affordable instead of replaying the entire transcript every turn.
3. Write in English when you can. Because of the ~1 token/character penalty, a Chinese prompt can cost 2–3× its English equivalent for the same content.
4. Chunk smarter in RAG. Retrieval quality and token cost trade off directly — measure chunk sizes in tokens, not characters.
Head back to the free token counter to count tokens with the exact o200k_base and cl100k_base encoders, see the per-model cost of your text, and check it against each model's context window. Everything runs in your browser — no account, no uploads.