# LLM Token Counter Estimate how many tokens your text will use — before you send it to a model or pay for it. This runs a BPE-aware approximation entirely in your browser: the real GPT pre-tokenisation step, plus a byte-calibrated subword model. No text is uploaded. ## What a token actually is Language models don't read characters or words — they read **tokens**, the units their tokeniser carves text into. A token is often a whole common word, sometimes a word fragment, sometimes a single punctuation mark or space. The rough rule of thumb is about **4 characters per token** for English, or three tokens for every four words — but that average hides a lot of variation, which is exactly what this tool tries to capture. ### How byte-pair encoding works Modern tokenisers use **byte-pair encoding (BPE)**. Training starts with raw bytes and repeatedly merges the most frequent adjacent pair into a new symbol, thousands of times over. The result is a vocabulary where common sequences ("the", " and", "ing") become single tokens, while rare strings stay split into smaller pieces. It's why "hello" is one token but a random string like "xqzkw" might be four. ### Why this is an approximation An exact count requires the model's full merge table — tens of thousands of learned rules, often a megabyte or more of data. Rather than ship that, this tool reproduces the part of the pipeline that does most of the work and needs no table: the **pre-tokenisation regex**. Real GPT tokenisers first split text with a pattern that isolates words (with their leading space), runs of punctuation, whitespace, and — importantly — numbers in groups of at most three digits. Each resulting piece is then estimated by its UTF-8 byte length at roughly four bytes per token. This captures the behaviours a naive "characters ÷ 4" misses entirely: numbers cost more than their length suggests, code and punctuation are token-dense, and non-Latin scripts like Chinese, Japanese, or Korean use several tokens per character because each character is multiple bytes. For ordinary English prose the estimate lands within a token or two of the real count; for code and mixed content it stays far closer than the simple rule. ### When you need the exact number For billing-critical or limit-critical work, confirm with the official tokeniser — OpenAI's `tiktoken`, or the count returned in an API response's usage field. Treat the number here as a fast, private, dependable *estimate* for drafting prompts, checking you're under a context window, and ballparking cost. ### Privacy Everything is computed locally in your browser. Your text is never transmitted, stored, or logged — paste a confidential prompt or document without concern. ## Accessibility ### Screen reader support --- *Source: [ocxly.com/token-counter.html](https://ocxly.com/token-counter.html) — OCXLY, free 100% client-side privacy-first tools. A free in-browser LLM token counter. Estimate token usage for GPT and other models with a BPE-aware approximation — real pre-tokenization plus a byte-calibrated subword model. From OCXLY AI Labs.*