AI Token Counter

Which models get an exact count

For Claude, Anthropic’s token counting endpoint is free and says its result is itself an estimate. It also says Claude 4.7 and later produce about 30% more tokens than earlier Claude models for the same text. For Gemini, Google’s token guide only offers “about 4 characters” per token and points to countTokens. OpenAI’s tiktoken has no entry for GPT-6, so this page does not claim one.

Examples

The sample prompt (the Sample button) is a 486-character code review request:

o200k_base gives 104 tokens, cl100k_base 104 and DeepSeek V4 105. With 1,000 output tokens, one request costs $0.03052 on gpt-5.5 and $0.001232 on deepseek-flash at peak rates (half that off-peak). For English the three tokenizers land within a few tokens of each other.

The gap opens with other scripts. The same one-line instruction in four languages:

The Chinese line costs 44% more tokens on GPT-4 (cl100k_base) than on GPT-4o (o200k_base), because the larger vocabulary holds many more Chinese words. Some characters are not in cl100k_base at all: 🦜 becomes byte tokens, which the token view shows in hex inside ⟨⟩ boxes.

Cost and context columns

Costs use the standard per-million-token prices checked on 2026-10-01, without batch or cache discounts. Where a provider charges more for long prompts, the table switches rate: gpt-5.5 and GPT-6 above 272,000 input tokens, gemini-3.1-pro-preview above 200,000. DeepSeek rows also show the off-peak price, which DeepSeek sets at half the peak price. The context column divides the count by each model’s input limit and marks rows over it.

How this differs from other counters

Tested on 2026-10-01 with the Chinese sentence “你是一名资深后端工程师，正在评审一个合并请求：在 PostgreSQL 分析库前面加一层 Redis 读穿缓存。” (o200k_base 35 tokens, DeepSeek V4 26):

• benchlm.ai token counter shows “~14 tokens” and lists 14 for every Claude, Gemini and DeepSeek model, about the character count divided by 4. The real counts are 35 and 26.
• tokencost.app marks DeepSeek V4.1 Flash as “42~” and lists GPT-6 Luna and Llama 3.3 with plain numbers, although OpenAI has published no GPT-6 tokenizer and Llama 3 uses its own vocabulary. Its DeepSeek price is the off-peak rate.
• token-counter.dev gives the correct 35 for o200k_base through tiktoken’s WebAssembly build, and has no DeepSeek tokenizer.

This page also fixes two problems of the common JavaScript port, js-tiktoken, which it used before. js-tiktoken reads \s with JavaScript’s rules, so text containing a byte-order mark (U+FEFF) or U+0085 can split differently from tiktoken. And its merge step slows down sharply on long runs without spaces: 20,000 Chinese characters in a row took over three minutes. Here the regex uses Unicode White_Space and the merge uses a priority queue, so the same input takes milliseconds.

Limits

• Counts are for plain text. Message wrappers, system prompts, tool schemas, images and PDFs add tokens that only the provider’s counting endpoint sees.
• Special tokens such as <|endoftext|> count as ordinary text for OpenAI (7 tokens), the same as tiktoken’s encode_ordinary; the page never throws on them, which the old version did. DeepSeek’s tokenizer.json recognises its own markers such as <｜User｜> as one token, and so does this page.
• Claude, Gemini and GPT-6 rows are references, not counts. Use the provider’s endpoint before you rely on them.
• Prices are a snapshot from 2026-10-01 and include no taxes, regional uplifts or negotiated discounts.
• DeepSeek’s package says to run it with transformers. In our test, transformers 5.18.0 returned an empty list for Chinese text from this package; transformers 4.57.6 and the tokenizers library give the ids this page uses.

Related: check plain word and character counts with the Word & Character Counter, remove API keys from a prompt with Redact API Keys & Secrets, and convert batch request files with the JSONL Converter.