Tiktoken vs Approximate Token Counts - Browser Guide
Last updated: July 2026
When browser counts match OpenAI tiktoken encodings (o200k, cl100k) and when Claude or Gemini stays approximate. How to read the labels and verify on the API.
Not every vendor ships a browser-ready tokenizer you can drop into a static page. LLM TokenBase is honest about that split: some modes aim for exact encodings; others are approximate and labeled that way.
What tiktoken-style modes can do
For OpenAI-oriented workflows we use encodings such as o200k_base and cl100k_base when they are available in the browser. That is the closest public match to how GPT-family APIs commonly talk about length. Open the token counter and switch to an OpenAI-style mode.
What we cannot ship for every vendor
Anthropic and Google do not give us the same static-page encoding twin. Claude and Gemini modes on the same counter still give a practical estimate — with the approximate badge kept visible. See the Claude and Gemini guides for more.
How to read the labels
Exact mode names follow encodings we ship (for example “OpenAI (o200k)”), not marketing nicknames. Newer IDs without a browser mapping stay in the methodology notes. Chat wrappers, tools, and multi-message formats can still change real usage.
Verify on the API when it matters
Planning counts belong on this site. Hard limits and invoices belong on the provider console.