LLM token counter

For English prose on OpenAI's o200k_base tokenizer (used by GPT-4o, GPT-4o mini, o-series, and GPT-5.x on that encoding), 1,000 words is often roughly 1,300 to 1,400 tokens. Code, Markdown tables, and heavy punctuation usually raise that ratio because symbols and rare strings split into more tokens.

Providers charge per million input tokens and per million output tokens. The pricing table multiplies your pasted token count by each model's listed rates for input, cached input, and output. For a full call, add (prompt tokens ÷ 1,000,000 × input or cached rate) plus (completion tokens ÷ 1,000,000 × output rate). The output columns show cost only if your entire paste were output tokens; estimate completion length separately. System prompts and RAG context count on every call if you resend them.

Counts are exact for models that use OpenAI's o200k_base encoding. They are not exact for models on other tokenizers, including older OpenAI cl100k_base families and many non-OpenAI providers. For invoice-level validation, use the vendor's official counter.

Rates and context windows are synced from OpenRouter into a local catalog and shown in the table. They are estimates for comparing models, not guaranteed invoice amounts. Providers change pricing and limits without notice; confirm before you budget or ship. The sync date appears below the table when available.

A red row means your current token count exceeds that model's listed context window in our catalog. The API may reject the request, truncate input, or drop older turns depending on the provider. Shrink the prompt, split the job, or pick a model with a larger context limit.

This page tokenizes with o200k_base (OpenAI's encoding). Claude and Gemini use their own tokenizers, so the same text can produce a different token count on their APIs. The table multiplies your o200k_base count by OpenRouter-listed rates for those models, which is useful for rough comparison but not a substitute for each vendor's official token counter when you need exact billing.

Select the Claude Fable 5 row in the pricing table and apply your prompt token count to input, cached-input, and output columns. This gives a planning estimate for Fable calls, but because Anthropic uses its own tokenizer, validate final billing-critical numbers with Anthropic's official token counting flow before launch.

API usage may include extra tokens beyond raw text, such as tool schemas, structured output constraints, images, files, and provider-specific wrappers. This tool is best for prompt-size planning; use provider token-count endpoints for full payload accounting.

Tokenizers are trained heavily on natural language. Source code packs in braces, operators, indentation, and rare identifiers that often split into several sub-tokens each. A line of code can cost noticeably more tokens than a line of prose; paste the real file here instead of guessing from word count.

Often yes. Many tokenizers assign fewer single-token words to non-Latin scripts (for example Arabic, Hindi, Japanese, or Chinese), so one visible word may become several tokens. The multiplier varies by language and text. Paste translated or multilingual copy here to measure o200k_base size for OpenAI prompts.

A context window is the maximum tokens one API call can include: system prompt, history, retrieved chunks, tools, and the new user message. Over the limit you may see errors, truncated replies, or dropped older turns. Very long prompts can also hurt recall in the middle of the context on some models. Count everything you plan to send, not just the latest message.

Prompt caching lets providers reuse identical prefix text (large system prompts, docs, or code summaries) on later requests at a lower cached rate when your payload structure qualifies. Providers define minimum sizes and layout rules. Use this counter to measure those static blocks, then compare the Input and Input (cached) columns in the table to see estimated cost for uncached versus cached input at your token count.

Yes. Tokenization runs in your browser with js-tiktoken; your text is not uploaded, logged, or stored on our servers. You can size internal system prompts, customer data, or code samples without sending them to us.

Characters are letters, digits, and symbols. Words are human units split on whitespace. Tokens are model-specific BPE pieces; one word may be one token or many. This page counts o200k_base tokens for OpenAI API sizing and shows words and characters alongside. For sentences, paragraphs, and reading time, use the word counter on this site.