The token counter that actually counts
Most token calculators divide your character count by four and print a model name next to it. This one runs real tokenizers and provider counting endpoints, so the number you see is the number you get billed for.
Pick your provider
Each counter uses the most exact method that exists for its provider. That method differs, and the difference matters.
Claude
Anthropic publishes no tokenizer. Exact counts come from their count_tokens API, per model, including the generation split.
OpenAI
o200k_base is open source. Counting runs in your browser and your text never leaves the page.
Gemini
Google's countTokens endpoint returns the exact number the API bills.
Kimi
Moonshot's tokenizer is open, with a 163,840-token vocabulary. Exact and in-browser.
DeepSeek
Open tokenizer, exact in-browser counting.
Tokens to words
Quick conversions both directions, including the 1M-token preset.
Why exact beats estimated
A characters-divided-by-four estimate can miss by 30% or more on code, non-English text, and structured data. At API prices that error compounds: budget projections, context-window planning, and prompt-size limits all inherit it. There is also no single answer to "how many tokens is this text": every provider tokenizes differently, and Claude tokenizes differently across model generations. A number without a named tokenizer behind it is a guess.
Our methods are public. Every counter names its mechanism, and we verify counts against provider APIs on a fixed 100-prompt corpus. Read how we count.
Beyond counting
API pricing
Every provider's list prices in one table, with a cost calculator that models cached input.
Context windows
Current limits for every major model, plus a fit checker for your prompt size.
What is a token?
Tokenization in plain language, and why counts differ between models.
VRAM calculator
Thinking of self-hosting instead? What your GPU can actually run.