LLM Token Counter

How to use LLM Token Counter

1. 1

   Paste text to count

   Paste any content: prompt text, article, code snippet, or conversation. OpenAI GPT and o-series rows are counted with the real o200k_base tokenizer, so those numbers are exact. Claude, Gemini, Grok, DeepSeek and Mistral publish no browser tokenizer, so their rows are character-ratio estimates.

2. 2

   Compare the per-family counts

   The same text is counted for every tokenizer family at once, so you can see how far the estimates diverge. There is no model to pick: all families are shown side by side.

3. 3

   See token estimate and cost breakdown

   Get total token count, approximate cost at current pricing, and cost per 1M tokens. Useful for budgeting API calls.

4. 4

   Load a file instead of pasting

   Upload a text file to count it without pasting. To compare several prompts, count them one at a time and note each result; there is no multi-prompt batch mode.

5. 5

   Test text reduction strategies

   Use the counter to compare token usage before and after shortening. Remove redundant words or summarize to optimize API costs.

Questions and answers

REST API

curl "https://aidevhub.io/api/token-counter/?text=Hello+world"

const res = await fetch("https://aidevhub.io/api/token-counter/", {
  method: "GET",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({
  "text": "Hello, world!"
}),
});
const data = await res.json();
console.log(data);

import requests

res = requests.post(
    "https://aidevhub.io/api/token-counter/",
    json={"text":"Hello, world!"},
)
print(res.json())

{
  "text_length": 13,
  "tokens": {
    "GPT-6 Astra": {
      "model": "GPT-6 Astra",
      "tokens": 4,
      "method": "exact",
      "cost": {
        "input": 0.000039999999999999996,
        "output": 0.00019999999999999998
      }
    }
  },
  "model_count": 39
}

How do I count tokens for GPT and Claude prompts with an offline tokenizer?

Paste your prompt into the counter above. It computes character and word counts, then estimates tokens per model family using the characters-per-token ratios below — all in the browser, with no API call and no network dependency after the page loads. This matters when prompts contain confidential material, or when you need counts inside an offline or air-gapped workflow.

Characters-per-token ratios by model family

Ratios apply to typical English prose; data as of October 3, 2026. A practical consequence: the same prompt usually produces more tokens on Claude (3.5 chars/token) than on GPT-family models (4.0 chars/token).

How accurate is a ratio-based token estimate?

A characters-per-token ratio is an estimate, not a byte-pair-encoding run. It is closest for plain English text and drifts for source code, dense punctuation, non-Latin scripts, and unusual whitespace, where real tokenizers split text differently. Use ratio estimates for sizing prompts, comparing models, and budgeting; use the provider's own tokenizer when you need billing-exact counts.

When do I need an exact count instead?

1. Enforcing hard context-window limits, where an overshoot causes a request to be rejected or silently truncated.
2. Metering or invoicing customers on token usage.
3. For OpenAI models, run the open-source tiktoken library locally; for Anthropic models, call the count-tokens endpoint of the API.
4. For everything before that point — drafting, comparing, budgeting — a client-side estimate is faster and keeps the prompt on your machine.