Prompt anatomy, decoded

See What Your Prompt Is Really Made Of.

Watch text turn into the exact raw-text tokens used by different GPT model families. Your prompt never leaves this browser.

01 / Input

Your Prompt

Paste or write any prompt. Encoding happens on this device.

02 / ModelChoose a tokenizer

Every model family ships its own vocabulary, so the same prompt splits differently — and costs a different number of tokens — depending on which one encodes it.

GPT-5 / 4oo200k_baseGPT-5, GPT-4.1, GPT-4o and the o-series reasoning models. ~200k vocabulary — the fewest tokens for most text.GPT-4 / 3.5cl100k_baseGPT-4, GPT-4 Turbo, GPT-3.5 Turbo and the text-embedding models. ~100k vocabulary.GPT-2r50k_baseGPT-2 and the original GPT-3 models. ~50k vocabulary — splits words and non-English text far more.

03 / Breakdown

Token Map

o200k_base

Tokens: —
Characters: —
UTF-8 bytes: —
Chars / token: —

How to Read This

Tokens Are Pieces, Not Words.

Each band is 1 token. Spaces and line breaks appear as · and ↵; byte fragments appear as \xNN. Color helps separate neighbors, while every band also carries its text and UTF-8 byte count. Select a band for its token ID.

This is the exact local encoding of the raw text only—not a billable request total. Messages, tools, images, system framing, and reasoning can add more tokens.

Claude and Gemini are intentionally absent because their exact production tokenizers are not available here for trustworthy local visualization. Orc does not present cross-provider guesses as facts.