Click anywhere in the text, or on a token ID, to inspect that token.
Models
Tokens
49
Exact for GPT-5 models (o200k_base). Each model is listed below.
- Characters
- 215
- Words
- 36
- Lines
- 3
- Chars / token
- 4.39
Where your tokens go
- Words78%
- Numbers2%
- Punctuation14%
- Whitespace2%
- Split characters4%
- 1 character takes more than one token (2 in all): emoji and letters the vocabulary rarely sees.
Context window by model
GPT-5.5
49
<0.1% of the 1.05M context window
GPT-6.1 Sol
≈ 49
<0.1% of the 1.05M context window
Claude Opus 5.5
≈ 86
<0.1% of the 1M context window
Claude Sonnet 5.5
≈ 86
<0.1% of the 1M context window
Gemini 4 Argon
≈ 54
<0.1% of the 1.05M context window
Grok 4.7
≈ 49
<0.1% of the 500K context window
DeepSeek V4 Pro
≈ 65
<0.1% of the 1M context window
How each count is made
- GPT-5.5ExactCounted with o200k_base, the vocabulary OpenAI's tiktoken uses for every GPT-5 model.
- GPT-6EstimateOpenAI hasn't published GPT-6's tokenizer, so this is counted with o200k_base, the vocabulary every GPT-5 model uses.
- ClaudeEstimateAnthropic doesn't publish Claude's tokenizer. Its docs say the current one makes about 2.5 characters per token (1M tokens is roughly 2.5M characters), so this is characters divided by 2.5.
- GeminiEstimateGoogle's docs say a Gemini token is about 4 characters, so this is characters divided by 4.
- GrokEstimatexAI doesn't publish Grok's tokenizer or a ratio, so this is counted with OpenAI's o200k_base as a stand-in.
- DeepSeekEstimateFrom DeepSeek's published rates: about 0.3 tokens per English character and 0.6 per Chinese character.
Context windows from each provider's model page, checked on 4 November 2026: OpenAI, Anthropic, Google, xAI, DeepSeek.
How this one works
The text is split with o200k_base, the tokenizer OpenAI publishes for its GPT-5 models, running in a background thread of this tab. Each coloured block in the field is one token. Put the cursor anywhere in the text, or click a token ID, to see that token's text with its spaces shown, its ID, its size in bytes and its position.
The breakdown sorts the tokens into words, numbers, punctuation, whitespace and characters the vocabulary had to split, such as emoji, and says when trimming whitespace would save tokens.
Most providers don't publish their tokenizers, so other models are estimated. For Claude, Gemini and DeepSeek the estimate uses the characters-per-token figure each provider documents; GPT-6 and Grok use o200k_base as a stand-in. The list under the models says which is which, and every estimate shows with ≈.
Questions
What is a token?
The unit a language model reads and writes: a common word, part of a longer word, a punctuation mark or a few bytes of an emoji. Models are priced and limited in tokens, not characters.
Why is my Claude count higher than my GPT count?
Anthropic's current tokenizer, used since Claude Opus 4.7, makes about 2.5 characters per token. OpenAI's makes about four for English. The same prompt is therefore roughly 60% more Claude tokens.
Why does a space get its own colour, or stick to the next word?
Tokenizers attach a leading space to the word after it, so " the" is one token. A run of spaces or a blank line usually becomes tokens of its own, which is why indentation in a prompt costs tokens. Show spaces makes them visible.
How do I get an exact count for Claude or Gemini?
Both providers have a token-counting API that returns the exact number for a request. It needs your API key, which is why it isn't used here.
Is my prompt sent anywhere?
No. The tokenizer runs in your browser and is downloaded once, the first time you edit the text. The prompt stays in memory and is gone when you close the tab.