LLM Token Calculator
Exact OpenAI token counts (tiktoken) plus $ estimates from published rates. Nothing uploaded.
Tokens exact (tiktoken) · $ estimates as of 2026-08-11
Mode
Model
Output tokens
OpenAI only. Claude/Gemini omitted — different tokenizers. Verify $ on OpenAI pricing before budgeting. Nothing uploaded.
Sponsored
Ad space
How to count LLM tokens
- Paste plain text, or switch to Chat messages and paste a JSON array.
- Pick the OpenAI model (encoding is shown in the selector).
- Optional: set expected output tokens to estimate completion cost.
- Read input tokens, context %, and $/request — compare models in the table.
Exact counts, honest costs
In 2026–2028 the hard part is not “how many words” — it is which tokenizer and which price row. This tool runs OpenAI's BPE encodings locally so GPT-4o / GPT-4.1 / GPT-5 / o-series / embedding prompts get exact token counts. Dollar figures multiply those counts by published per-million rates (last aligned 2026-08-11) — always recheck OpenAI before budgeting. We do not pretend Claude or Gemini are tiktoken.
Need a checksum instead of tokens? Try the Hash Generator. Encoding payloads for APIs? Use Base64 Encode / Decode.
FAQ
Are token counts exact?
Yes for the OpenAI models listed. Counting uses official tiktoken encodings (o200k_base / cl100k_base) via gpt-tokenizer in your browser.
Are the dollar amounts exact?
No — they are estimates. Exact tokens × published $/1M rates (table last aligned 2026-08-11). OpenAI changes prices; confirm on the official pricing page before budgeting.
Why isn’t Claude or Gemini included?
They use different tokenizers. Using OpenAI tiktoken for them is often wrong by 10–20%+. We only show counts we can make exact client-side.
What does Chat messages mode do?
It counts a JSON array of { role, content } with ChatML overhead (encodeChat) for the selected model — closer to chat/completions prompt billing than raw string length.
Why can output cost dominate?
You enter assumed completion tokens. Reasoning models (o3, o4-mini, …) also bill hidden reasoning tokens as output, so real bills can exceed the visible reply.
Is my prompt uploaded?
No. Tokenization runs entirely in your browser.
Sponsored
Ad space